{
  "File Number": "100",
  "Title": "Invariance Principle Meets Information Bottleneck for Out-of-Distribution Generalization",
  "6 Extensions, limitations, and future work": "Extension to non-linear models and multi-class classification. In this work our theoretical analysis focused on linear models. Consider the map X ← S(Zinv, Zspu) in Assumption 2. Suppose S is non-linear and bijective. We can divide the learning task into two parts a) invert S to obtain Zinv, Zspu and b) learn a linear model that only relies on the invariant features Zinv to predict the label Y . For\n9We place Terra Incognita and COCO dataset in the FIIF assuming that the humans who labeled the images did not need to rely on unreliable/spurious features such as background to generate the labels.\npart b), we can rely on the approaches proposed in this work. For part a), we need to leverage advancements in the field of non-linear ICA (Khemakhem et al., 2020). The current state-of-the-art to solve part a) requires strong structural assumptions on the dependence between all the components of Zinv, Zspu (Lu et al., 2021). Therefore, solving part a) and part b) in conjunction with minimal assumptions forms an exciting future work. In the entire work, the discussion was focused on binary classification tasks and regression tasks. For multi-class classification settings, we consider natural extension of the SEM in Assumption 2 (See the Appendix) and our main results continue to hold.\nOn the choice for IB penalty and IRMv1 penalty. We use the approximation for entropy (in equation (7)) described in Kirsch et al. (2020). The approximation (even though an upper bound) serves as an effective proxy for the true information bottleneck as shown in the experiments in Kirsch et al. (2020) (e.g., see their experiment on Imagenette dataset). Also, our experiments validate this approximation even in moderately high dimensions, as an example in CS-CMNIST, the dimension of the layer at which bottleneck constraints are applied is 256. Developing tighter approximations for information bottleneck in high dimensions and analyzing their impact on OOD generalization is an important future work. In recent works (Rosenfeld et al., 2021; Kamath et al., 2021; Gulrajani and Lopez-Paz, 2021), there has been criticism of different aspects of IRM, e.g., failure of IRMv1 penalty in non-linear models, the tuning of IRMv1 penalty, etc. Since we use IRMv1 penalty in our proposed loss, these criticisms apply to our objective as well. Other approximations of invariance have been proposed in the literature (Koyama and Yamaguchi, 2020; Ahuja et al., 2020; Chang et al., 2020). Exploring their benefits together with information bottleneck is a fruitful future work. Before concluding, we want to remark that we have already discussed the closest related works. However, we also provide a detailed discussion of the broader related literature in the Appendix.",
  "Reviewer Comment": "Reviewer_1: I pretty like the theoretical analysis presented in this paper, from which we could get very insightful thoughts.\nIn this paper, the authors use 0-1 loss for binary classification\nY\n=\n0\n,\n1\nand square loss for regression\nY\n=\nR\n. I am curious if all the results in this paper can be extended to the other losses and to multi-class classification tasks.\nAlthough the authors admit in the final section that the proposed method mainly focuses on linear models. I would like to hear more from the authors about what the main challenge is when their theory is extended to nonlinear models.\nLimitations And Societal Impact:\nYes.\nNeeds Ethics Review: No\nTime Spent Reviewing: two hours\n\nReviewer_2: In general, this work made a few theoretical contributions. It specifies some conditions to guarantee good generalization of IRM in classification, especially the support overlap of invariant (or spurious) features. Some theorems (like Theorem 4) also partially explains why information bottleneck can improve generalization, which is less investigated before.\nHowever, I still have some concerns:\nAuthors combine the objective of IRM and the minimizing of entropy of representation\nH\n(\nΦ\n(\nX\n)\n)\n. However, instead of directly optimizing the differential entropy (which is usually hard), authors made a Gaussian assumption and alternatively minimize the variance. I am not so sure how practical is this assumption in real-world data or applications. Because the experiments in this work is just on simple synthetic data or MNIST-like data.\nCan you provide more justifications on this approximation or more empricial results on recent domain generalization benchmarks (i.e., the DomainBed [1] or WILDS [2]) in which both IRM and ERM are baselines.\n[1] Gulrajani, Ishaan, and David Lopez-Paz. \"In search of lost domain generalization.\" arXiv preprint arXiv:2007.01434 (2020).\n[2] Koh, Pang Wei, et al. \"Wilds: A benchmark of in-the-wild distribution shifts.\" International Conference on Machine Learning. PMLR, 2021.\nSimilar to the information bottleneck, I feel the hyperparameter\nγ\n(in Eq. (7)) controls a trade-off and plays a significant role to the performance. Here, there is another hyperprameter\nλ\nthat controls the degree of invariance. Can you give a more indepth analysis on the trade-offs between these terms or how different values of\nγ\nor\nλ\ninfluence the performance of the combined objective?\nI still feel there is a jump from your analysis in Section 3 to the introduction of information bottleneck regularization (i.e.,\nmin\nH\n(\nΦ\n(\nX\n)\n)\n) in Section 4. Just focusing on your illustrative example, when\nΦ\nonly picks invariant features, the entropy is minimum. However, it does not mean when the entropy is minimized,\nΦ\nexactly picks the invariant features. For example, the suprious features only also gives the minimum entropy. My question is how can you guarantee this regularization alone is sufficient for IB-IRM to only focus on invariant features, rather than other combinations of invariant and suprious features? or did I miss something?\nFinally, some minor suggestions:\nIt would be much better if authors can give some illustrative examples on \"fully\" informative features and \"partially\" informative features.\nLimitations And Societal Impact:\nAuthors clearly mentioned limitations and future work directions.\nNeeds Ethics Review: No\nTime Spent Reviewing: 6\n\nReviewer_3: This paper is well organized and easy to follow. The theoretical results are presented in a logical order and the conclusions are quite clear. Although all the proofs are in the appendix, authors provide the intuitions by examples and discussed the significance of the theorems. With these theorems, the proposed IB-ERM is well-motivated. Some relaxations are involved from the original formulation in (6) to the final objective in (7). The influence of the IB term to optimization is discussed in Theorem 5.\nOverall, this paper is well-written, with comprehensive theoretical analysis to support the proposed IB-ERM for linear classification. Detailed comments/questions: 1. Although we can construct tasks like CS-CMNIST and the model in assumption 2, is it true that, with high-probability, the oracle in real world is PIIF? Since the conditional independency is a special case of the conditional dependency. Then, what is the advantage of IB-IRM, comparing to other methods designed for PIIF, e.g., (Rosenfeld et al. 2021)? 2. The proposed IB-IRM in eqn (6) takes IB as the objective and ERM and invariance as constraints. Will theorem 4 still holds if one take IB as an additional constraint in IRM? Can we have more insights to design the relaxed objective later? 3. How did you tune the coefficient of the two penalty terms at the equation (7) in real experiments? Any ablation study to show the influence of these two penalties in a different situation? 4. In sec 5, authors make discussion about the results on CMNIST tasks. How about the experimental results on linear unit tests? Why only make comparison on three and six training environments? Do we have any prior knowledge about which tasks are PIIF and which tasks are FIIP? IB-IRM is not the best across all the tasks. 5. Minors: 1) put the DAG into Figure 2? 2) ref the definition of “support overlap invariant/spurious features” in table 1; 3) Is the observation in figure 3 general for different values of \\lambda and \\gamma? 4) In Theorem 3, “some of the ERM and IRM solution..” is not rigorous. “these exist”?\nLimitations And Societal Impact:\nSee the detailed questions in the main review, especially question 3 and 4.\nNeeds Ethics Review: No\nTime Spent Reviewing: 6\n\nReviewer_4: Strengths:\nThey argue that the information bottleneck in some situations can help to obtain the invariant predictor when invariance itself cannot (when there’s not enough support overlap)\nIn their (few) experiments on modified colored MNIST datasets, including one where IRM does not improve much compared to ERM (CS-MNIST) and demonstrate that adding the IB regularizer leads to better OOD generalization\nWeaknesses:\nTheory\nThe insufficiency results for IRM are not surprising, e.g. FIIF problem was already noted in Nagarajan et al\nComparison with other works that show limits of IRM such as Srebro - their examples themselves seem to make a similar point as FIIF?\nSince the theoretical contributions are insufficient, we now examine the methodological contributions Methodology:\nmajor criticism: is this supposed to be a game-changer for practice? then it’s absolutely necessary to do OOD generalization experiments on real world datasets (such as WILDS). the current datasets are completely toy (versions of colored MNIST)\nsince variances are huge compared to MSE for examples 1/1s, 3/3s with 6 environments, one can really only compare the performances of the methods using CS-MNIST and AC-MNIST and perhaps example 2/2s which is rather unsatisfying as empirical evidence for a method\ntable 3 suggests that more environments worsens the benefits of adding IB even in an FIIF setting? more discussion/experiments on that would be necessary\nagain regarding the point that the variances are huge for Example 1/1s with 6 vs. 3 environments: a comment on that in the main text would be useful - is that expected and why?\nUpdate: Apologies, I actually was not able to understand the problem that I previously had when I had a brief new look at the paper (before even reading your response). I'm guessing I was thinking about some problems that confused my mindset when reading. Now the sufficiency result actually sounded trivial since in the case of spurious correlations AND support overlap all possible environments in E_all would also be spuriously correlated...\nThanks for discussion on relation to Kamath et al.\nI agree Nagarajan did not provide formal results, but hinted at the problem so this is why I said it is not surprising.\nThanks for comment on six-environments to discuss the high variance. That comment should be added to paper.\nAdding experiments on Coco addresses my main concern (I still think the methodology is the main contribution). However I would like to see more experimental details before giving the nod + experiments on one more dataset. The lack of real-world experimental evidence was the primal reason for the score and hence I still tend to reject the paper. The number of scenarios where IB-IRM or IB-ERM do better is still rather limited and currently only includes COCO in terms of non-toy dataset (without having been able to see the revision and experimental details) which is in my opinion still quite rather meager for publication at Neurips. How about WILDS? I believe if the authors run a more extensive experimental study this paper can be accepted at the next venue.\nI didn't see this point addressed: \"Table 3 suggests that more environments worsens the benefits of adding IB even in an FIIF setting? more discussion/experiments on that would be necessary\"\nLimitations And Societal Impact:\nAuthors do have a section on limitations and motivate future work through that. There's no potential negative societal impact for OOD detection per se ... (for certain applications it could, but this holds for almost all problems that one can think of solving).\nNeeds Ethics Review: No\nTime Spent Reviewing: 7",
  "abstractText": "The invariance principle from causality is at the heart of notable approaches such as invariant risk minimization (IRM) that seek to address out-of-distribution (OOD) generalization failures. Despite the promising theory, invariance principle-based approaches fail in common classification tasks, where invariant (causal) features capture all the information about the label. Are these failures due to the methods failing to capture the invariance? Or is the invariance principle itself insufficient? To answer these questions, we revisit the fundamental assumptions in linear regression tasks, where invariance-based approaches were shown to provably generalize OOD. In contrast to the linear regression tasks, we show that for linear classification tasks we need much stronger restrictions on the distribution shifts, or otherwise OOD generalization is impossible. Furthermore, even with appropriate restrictions on distribution shifts in place, we show that the invariance principle alone is insufficient. We prove that a form of the information bottleneck constraint along with invariance helps address key failures when invariant features capture all the information about the label and also retains the existing success when they do not. We propose an approach that incorporates both of these principles and demonstrate its effectiveness in several experiments.",
  "1 Introduction": "Recent years have witnessed an explosion of examples showing deep learning models are prone to exploiting shortcuts (spurious features) (Geirhos et al., 2020; Pezeshki et al., 2020) which make them fail to generalize out-of-distribution (OOD). In Beery et al. (2018), a convolutional neural network was trained to classify camels from cows; however, it was found that the model relied on the background color (e.g., green pastures for cows) and not on the properties of the animals (e.g., shape). These examples become very concerning when they occur in real-life applications (e.g., COVID-19 detection (DeGrave et al., 2020)).\nTo address these out-of-distribution generalization failures, invariant risk minimization (Arjovsky et al., 2019) and several other works were proposed (Ahuja et al., 2020; Pezeshki et al., 2020; Krueger et al., 2020; Robey et al., 2021; Zhang et al., 2021). The invariance principle from causality (Peters et al., 2015; Pearl, 1995) is at the heart of these works. The principle distinguishes predictors that only rely on the causes of the label from those that do not. The optimal predictor that only focuses on the causes is invariant and min-max optimal (Rojas-Carulla et al., 2018; Koyama and Yamaguchi, 2020; Ahuja et al., 2021) under many distribution shifts but the same is not true for other predictors. ∗Equal contribution. †Mila - Quebec AI Institute, Université de Montréal. Correspondence to: kartik.ahuja@mila.quebec.\n35th Conference on Neural Information Processing Systems (NeurIPS 2021).\nOur contributions. Despite the promising theory, invariance principle-based approaches fail in settings (Aubin et al., 2021) where invariant features capture all information about the label contained in the input. A particular example is image classification (e.g., cow vs. camel) (Beery et al., 2018) where the label is a deterministic function of the invariant features (e.g., shape of the animal), and does not depend on the spurious features (e.g., background). To understand such failures, we revisit the fundamental assumptions in linear regression tasks, where invariance-based approaches were shown to provably generalize OOD. We show that, in contrast to the linear regression tasks, OOD generalization is significantly harder for linear classification tasks; we need much stronger restrictions in the form of support overlap assumptions3 on the distribution shifts, or otherwise it is not possible to guarantee OOD generalization under interventions on variables other than the target class. We then proceed to show that, even under the right assumptions on distribution shifts, the invariance principle is insufficient. However, we establish that information bottleneck (IB) constraints (Tishby et al., 2000), together with the invariance principle, provably works in both settings – when invariant features completely capture the information about the label and also when they do not. (Table 1 summarizes our theoretical results presented later). We propose an approach that combines both these principles and demonstrate its effectiveness on linear unit tests (Aubin et al., 2021) and on different real datasets.",
  "2 OOD generalization and invariance: background & failures": "Background. We consider a supervised training data D gathered from a set of training environments Etr: D = {De}e∈Etr , where De = {xei , yei }n e\ni=1 is the dataset from environment e ∈ Etr and ne is the number of instances in environment e. xei ∈ Rd and yei ∈ Y ⊆ Rk correspond to the input feature value and the label for ith instance respectively. Each (xei , y e i ) is an i.i.d. draw from Pe, where Pe is the joint distribution of the input feature and the label in environment e. Let X e be the support of the input feature values in the environment e. The goal of OOD generalization is to use training data D to construct a predictor f : Rd → Rk that performs well across many unseen environments in Eall, where Eall ⊃ Etr. Define the risk of f in environment e as Re(f) = E [ `(f(Xe), Y e) ] , where for example ` can be 0-1 loss, logistic loss, square loss, (Xe, Y e) ∼ Pe, and the expectation E is w.r.t. Pe. Formally stated, our goal is to use the data from training environments Etr to find f : Rd → Y to minimize\nmin f max e∈Eall\nRe(f). (1)\nSo far we did not state any restrictions on Eall. Consider binary classification: without any restrictions on Eall, no method can reduce the above objective (` is 0-1 loss) to below one. Suppose a method outputs f∗; if ∃ e ∈ Eall \\ Etr with labels based on 1− f∗, then it achieves an error of one. Some assumptions on Eall are thus necessary. Consider how Eall is restricted using invariance for linear regressions (Arjovsky et al., 2019). Assumption 1. Linear regression structural equation model (SEM). In each e ∈ Eall\nY e ← w∗inv · Zeinv + e, Zeinv ⊥ e, E[ e] = 0,E [ | e|2 ] ≤ σ2sup Xe ← S(Zeinv, Zespu) (2)\nwhere w∗inv ∈ Rm, Zeinv ∈ Rm, Zspu ∈ Ro, S ∈ Rd×(m+o), S is invertible (m+ o = d). We focus on invertible S but several results extend to non-invertible S as well (see Appendix).\n3Support is the region where the probability density for continuous random variables (probability mass function for discrete random variables) is positive. Support overlap refers to the setting where train and test distribution maybe different but share the same support. We formally define this later in Assumption 5.\nAssumption 1 states how Y e andXe are generated from latent invariant features Zeinv 4, latent spurious features Zespu and noise e. The relationship between label and invariant features is invariant, i.e., w∗inv is fixed across all environments. However, the distributions of Z e inv, Z e spu, and\ne are allowed to change arbitrarily across all the environments. Suppose S is identity. If we regress only on the invariant features Zeinv, then the optimal solution is w ∗ inv, which is independent of the environment, and the error it achieves is bounded above by the variance of e (σ2sup). If we regress on the entire Ze and the optimal predictor places a non-zero weight on Zespu (e.g., Z e spu ← Y e + ζe), then this predictor fails to solve equation (1) (∃ e ∈ Eall, Zespu →∞, error→∞, see Appendix for details). Also, not only regressing on Zeinv is better than on Z\ne, it can be shown that it is optimal, i.e., it solves equation (1) under Assumption 1 and achieves a value of σ2sup for the objective in equation (1).\nInvariant predictor. Define a linear representation map Φ : Rr×d (that transforms Xe as Φ(Xe)) and define a linear classifier w : Rk×r (that operates on the representation w · Φ(Xe)). We want to search for representations Φ such that E[Y e|Φ(Xe)] is invariant (in Assumption 1 if Φ(Xe) = Zeinv, then E[Y e|Φ(Xe)] is invariant). We say that a data representation Φ elicits an invariant predictor w ·Φ across the set of training environments Etr if there is a predictor w that simultaneously achieves the minimum risk, i.e., w ∈ arg minw̃ Re(w̃ · Φ), ∀e ∈ Etr. The main objective of IRM is stated as\nmin w∈Rk×r,Φ∈Rr×d\n1 |Etr| ∑ e∈Etr Re(w · Φ) s.t. w ∈ arg min w̃∈Rk×r Re(w̃ · Φ), ∀e ∈ Etr. (3)\nObserve that if we drop the constraints in the above which search only over invariant predictors, then we get the standard empirical risk minimization (ERM) (Vapnik, 1992) (assuming all the training environments occur with equal probability). In all our theorems, we use 0-1 loss for binary classification Y = {0, 1} and square loss for regression Y = R. For binary classification, the output of the predictor is given as I(w ·Φ(Xe)), where I(·) is the indicator function that takes 1 if the input is ≥ 0 and 0 otherwise, and the risk is Re(w ·Φ) = E [ |I(w ·Φ(Xe))−Y e| ] . For regression, the output\nof the predictor is w ·Φ(Xe) and the corresponding risk is Re(w ·Φ) = E [ (w ·Φ(Xe)− Y e)2 ] . We now present the main OOD generalization result from Arjovsky et al. (2019) for linear regressions.\nTheorem 1. (Informal) If Assumption 1 is satisfied, Rank[Φ] > 0, |Etr| > 2d, and Etr lie in a linear general position (a mild condition on the data in Etr, defined in the Appendix), then each solution to equation (3) achieves OOD generalization (solves equation (1), @ e ∈ Eall with risk > σ2sup).\nDespite the above guarantees, IRM has been shown to fail in several cases including linear SEMs in (Aubin et al., 2021). We take a closer look at these failures next.\nUnderstanding the failures: fully informative invariant features vs. partially informative invariant features (FIIF vs. PIIF). We define properties salient to the datasets/SEMs used in the OOD generalization literature. Each e ∈ Eall, the distribution (Xe, Y e) ∼ Pe satisfies the following properties. a) ∃ a map Φ∗ (linear or not), which we call an invariant feature map, such that E [ Y e ∣∣Φ∗(Xe)] is the same for all e ∈ Eall and Y e 6⊥ Φ∗(Xe). These conditions ensure Φ∗ maps to features that have a finite predictive power and have the same optimal predictor across Eall. For the SEM in Assumption 1, Φ∗ maps to Zeinv. b) ∃ a map Ψ∗ (linear or not), which we call spurious feature map, such that E [ Y e ∣∣Ψ∗(Xe)] is not the same for all e ∈ Eall and Y e 6⊥ Ψ∗(Xe) for some environments. Ψ∗ often creates a hindrance in learning predictors that only rely on Φ∗. Note that Ψ∗ should not be a transformation of some Φ∗. For the SEM in Assumption 1, suppose Zespu is anti-causally related to Y e, then Ψ∗ maps to Zespu (See Appendix for an example).\nIn the colored MNIST (CMNIST) dataset (Arjovsky et al., 2019), the digits are colored in such a way that in the training domain, color is highly predictive of the digit label but this correlation being spurious breaks down at test time. Suppose the invariant feature map Φ∗ extracts the uncolored digit and the spurious feature map Ψ∗ extracts the background color. Ahuja et al. (2021) studied two variations of the colored MNIST dataset, which differed in the way final labels are generated from original MNIST labels (corrupted with noise or not). They showed that the IRM exhibits good OOD generalization (50% improvement over ERM) in anti-causal-CMNIST (AC-CMNIST, original data from Arjovsky et al. (2019)) but is no different from ERM and fails in covariate shift-CMNIST (CSCMNIST). In AC-CMNIST, the invariant features Φ∗(Xe) (uncolored digit) are partially informative about the label, i.e., Y 6⊥ Xe|Φ∗(Xe), and color contains information about label not contained\n4In many examples in the literature, invariant features are causal, but not always (Rosenfeld et al., 2021).\nin the uncolored digit. On the other hand in CS-CMNIST, invariant features are fully informative about the label, i.e., Y ⊥ Xe|Φ∗(Xe), i.e., they contains all the information about the label that is contained in input Xe. Most human labelled datasets have fully informative invariant features; the labels (digit value) only depend on the invariant features (uncolored digit) and spurious features (color of the digit) do not affect the label. 5 In the rare case, when the humans are asked to label images in which the object being labelled itself is blurred, humans can rely on spurious features such as the background making such a data representative of PIIF setting. In Table 2, we divide the different datasets used in the literature based on informativeness of the invariant features. We observe that when the invariant features are fully informative, both IRM and ERM fail but only in classification tasks and not in regression tasks (Ahuja et al., 2021); this is consistent with the linear regression result in Theorem 1, where IRM succeeds regardless of whether Y e ⊥ Xe|Zeinv holds or not. Motivated by this observation, we take a closer look at the classification tasks where invariant features are fully informative.",
  "3 OOD generalization theory for linear classification tasks": "A two-dimensional example with fully informative invariant features. We start with a 2D classification example (based on Nagarajan et al. (2021)), which can be understood as a simplified version of the CS-CMNIST dataset (Ahuja et al., 2021), Example 2/2S of Aubin et al. (2021), where both IRM and ERM fail. The example goes as follows. In each training environment e ∈ Etr\nY e ← I ( Xeinv − 1\n2\n) , where Xeinv ∈ {0, 1} is Bernoulli (1 2 ) ,\nXespu ← Xeinv ⊕W e, where W e ∈ {0, 1} is Bernoulli ( 1− pe ) with selection bias pe > 1\n2 ,\n(4)\nwhere Bernoulli(a) takes value 1 with probability a and 0 otherwise. Each training environment is characterized by the probability pe. Following Assumption 1, we assume that the labelling function does not change from Etr to Eall, thus the relation between the label and the invariant features does not change. Assume that the distribution of Xeinv and X e spu can change arbitrarily. See Figure 1a) for a pictorial representation of this example illustrating the gist of the problem: there are many classifiers with the same error on Etr while only the one identical to the labelling function I(Xeinv− 12 ) generalizes correctly OOD. Define a classifier I(winvxinv +wspuxspu − 12 (winv +wspu)). Define a set of classifiers S = {(winv, wspu) s.t. winv > |wspu|}. Observe that all the classifiers in S achieve a zero classification error on the training environments. However, only classifiers for which wspu = 0 solve the OOD generalization (eq. (1)). With Φ as the identity, it can be shown that all the classifiers S form an invariant predictor (satisfy the constraint in equation (3) over all the training environments when ` is the 0-1 loss). Observe that increasing the number of training environments to infinity does not address the problem, unlike with the linear regression result discussed in Theorem 1 (Arjovsky et al., 2019), where it was shown that if the number of environments increases linearly in the dimension of the data, then the solution to IRM also solves the OOD generalization (eq. (1)). 6 We use the above example to construct general SEMs for linear classification when the invariant features are fully informative. We follow the structure of the SEM from Assumption 1 in our construction.\n5The deterministic labelling case was referred as realizable problems in (Arjovsky et al., 2019). 6Please note that this example illustrates certain important facets in a very simple fashion; only in this example a max-margin classifier can solve the problem but not in general. (Further explanation in the Appendix).\nAssumption 2. Linear classification structural equation model (FIIF). In each e ∈ Eall\nY e ← I ( w∗inv · Zeinv ) ⊕Ne, Ne ∼ Bernoulli(q), q < 1\n2 , Ne ⊥ (Zeinv, Zespu), Xe ← S ( Zeinv, Z e spu ) ,\n(5)\nwhere w∗inv ∈ Rm with ‖w∗inv‖ = 1 is the labelling hyperplane, Zeinv ∈ Rm, Zespu ∈ Ro, Ne is binary noise with identical distribution across environments, ⊕ is the XOR operator, S is invertible.\nIf noise level q is zero, then the above SEM covers linearly separable problems. See Figure 2a) for the directed acyclic graph (DAG) corresponding to this SEM. From the DAG observe that Y e ⊥ Xe|Zeinv, which implies that the invariant features are fully informative. Contrast this with a DAG that follows Assumption 1 shown in Figure 2b), where Y e 6⊥ Xe|Zeinv and thus the invariant features are not fully informative. If Eall follows the SEM in Assumption 2 and suppose the distribution of Zeinv, Zespu can change arbitrarily, then it can be shown that only a classifier identical to the labelling function I(w∗inv · Zeinv) can solve the OOD generalization (eq. (1)); such a classifier achieves an error of q (noise level) in all the environments. As a result, if for a classifier we can find e ∈ Eall that follows Assumption 2 where the error is greater than q, then such a classifier does not solve equation (1). Now we ask – what are the minimal conditions on training environments Etr to achieve OOD generalization when Eall follow Assumption 2? To achieve OOD generalization for linear regressions, in Theorem 1, it was required that the number of training environments grows linearly in the dimension of the data. However, there was no restriction on the support of the latent invariant and latent spurious features, and they were allowed to change arbitrarily from train to test (for further discussion on this, see the Appendix). Can we continue to work with similar assumptions for the SEM in Assumption 2 and solve the OOD generalization (eq. (1))? We state some assumptions and notations to answer that. Define the support of the invariant (spurious) features Zeinv (Z e spu) in environment e as Zeinv (Zespu).\nAssumption 3. Bounded invariant features. ∪e∈EtrZeinv is a bounded set.7 Assumption 4. Bounded spurious features. ∪e∈EtrZespu is a bounded set.\nAssumption 5. Invariant feature support overlap. ∀e ∈ Eall,Zeinv ⊆ ∪e′∈EtrZe ′ inv Assumption 6. Spurious feature support overlap. ∀e ∈ Eall,Zespu ⊆ ∪e′∈EtrZe ′ spu\nAssumption 5 (6) states that the support of the invariant (spurious) features for unseen environments is the same as the union of the support over the training environments. It is important to note that support overlap does not imply that the distribution over the invariant features does not change. We now define a margin that measures how much the is training support of invariant features Zeinv separated by the labelling hyperplane w∗inv. Define Inv-Margin = minz∈∪e∈EtrZeinv sgn ( w∗inv · z )( w∗inv · z ) . This margin only coincides with the standard margin in support vector machines when the noise level q is 0 (linearly separable) and S is identity. If Inv-Margin > 0, then the labelling hyperplane w∗inv separates the support into two halves (see Figure 1b)).\n7A set Z is bounded if ∃M <∞ such that ∀z ∈ Z, ‖z‖ ≤M .\nAssumption 7. Strictly separable invariant features. Inv-Margin > 0.\nNext, we show the importance of support overlap for invariant features.\nTheorem 2. Impossibility of guaranteed OOD generalization for linear classification. Suppose each e ∈ Eall follows Assumption 2. If for all the training environments Etr, the latent invariant features are bounded and strictly separable, i.e., Assumption 3 and 7 hold, then every deterministic algorithm fails to solve the OOD generalization (eq. (1)), i.e., for the output of every algorithm ∃ e ∈ Eall in which the error exceeds the minimum required value q (noise level).\nThe proofs to all the theorems are in the Appendix. We provide a high-level intuiton as to why invariant feature support overlap is crucial to the impossibility result. In Figure 1b), we show that if the support of latent invariant features are strictly separated by the labelling hyperplane w∗inv, then we can find another valid hyperplane w+inv that is equally likely to have generated the same data. There is no algorithm that can distinguish between w∗inv and w + inv. As a result, if we use data from the region where the hyperplanes disagree (yellow region Figure 1b)), then the algorithm fails.\nSignificance of Theorem 2. We showed that without the support overlap assumption on the invariant features, OOD generalization is impossible for linear classification tasks. This is in contrast to linear regression in Theorem 1 (Arjovsky et al., 2019), where even in the absence of the support overlap assumption, guaranteed OOD generalization was possible. Applying the above Theorem 2 to the 2D case (eq. (4)) implies that we cannot assume that the support of invariant latent features can change, or else that case is also impossible to solve.\nNext, we ask what further assumptions are minimally needed to be able to solve the OOD generalization (eq. (1)). Each classifier can be written as w̄ ·Xe = w̄ · S(Zeinv, Zespu) = w̃inv · Zeinv + w̃spuZespu. If w̃spu 6= 0, then the classifier w̄ is said to rely on spurious features. Theorem 3. Sufficiency and Insufficiency of ERM and IRM. Suppose each e ∈ Eall follows Assumption 2. Assume that a) the invariant features are strictly separable, bounded, and satisfy support overlap, b) the spurious features are bounded (Assumptions 3-5, 7 hold).\n• Sufficiency: If the spurious features satisfy support overlap (Assumption 6 holds), then both ERM and IRM solve the OOD generalization problem (eq. (1)). Also, there exist solutions to ERM and IRM solutions that rely on the spurious features and still achieve OOD generalization.\n• Insufficiency: If spurious features do not satisfy support overlap, then both ERM and IRM fail at solving the OOD generalization problem (eq. (1)). Also, there exist no such classifiers that rely on spurious features and also achieve OOD generalization.\nSignificance of Theorem 3. From the first part, we learn that if the support overlap is satisfied for both the invariant features and the spurious features, then either ERM or IRM can solve the OOD generalization (eq. (1)). Interestingly, in this case we can have classifiers that rely on the spurious features and yet solve the OOD generalization (eq. (1)). For the 2D case (eq. (4)) this case implies that the entire set S solves the OOD generalization (eq. (1)). From the second part, we learn that if support overlap holds for invariant features but not for spurious features, then the ideal OOD optimal predictors rely only on the invariant features. In this case, methods like ERM and IRM continue to rely on spurious features and fail at OOD generalization. For the above 2D case (eq. (4)) this implies that only the predictors that rely only on Xeinv in the set S solve the OOD generalization (eq. (1)). To summarize, we looked at SEMs for classification tasks when invariant features are fully informative, and find that the support overlap assumption over invariant features is necessary. Even in the presence of support overlap for invariant features, we showed that ERM and IRM can easily fail if the support overlap is violated for spurious features. This raises a natural question – Can we even solve the case with the support overlap assumption only on the invariant features? We will now show that the information bottleneck principle can help tackle these cases.",
  "4 Information bottleneck principle meets invariance principle": "Why the information bottleneck? The information bottleneck principle prescribes to learn a representation that compresses the input X as much as possible while preserving all the relevant information about the target label Y (Tishby et al., 2000). Mutual information I(X; Φ(X)) is used to measure information compression. If representation Φ(X) is a deterministic transformation of X , then in principle we can use the entropy of Φ(X) to measure compression (Kirsch et al., 2020). Let\nus revisit the 2D case (eq. (4)) and apply this principle to it. Following the second part of Theorem 3, where ERM and IRM failed, assume that invariant features satisfy the support overlap assumption, but make no such assumption for the spurious features. Consider three choices for Φ: identity (selects both features), selects invariant feature only, selects spurious feature only. The entropy of H(Φ(Xe)) when Φ is the identity is H(pe) + log(2), where H(pe) is the Shannon entropy in Bernoulli(pe). If Φ selects the invariant/spurious features only, then H(Φ(Xe)) = log(2). Among all three choices, the one that has the least entropy and also achieves zero error is the representation that focuses on the invariant feature. We could find the OOD optimal predictor in this example just by using information bottleneck. Does it mean the invariance principle isn’t needed? We answer this next.\nWhy invariance? Consider a simple classification SEM. In each e ∈ Etr, Y e ← X1,einv ⊕X 2,e inv ⊕Ne and Xespu ← Y e ⊕ V e, where all the random variables involved are binary valued, noise Ne, V e are Bernoulli with parameters q (identical across Etr), ce (varies across Etr) respectively. If ce < q, then in Etr predictions based on Xespu are better than predictions based on X 1,e inv , X 2,e inv . If both X1,einv , X 2,e inv are uniform Bernoulli, then these features have a higher entropy than X e spu. In this case, the information bottleneck would bar using X1,einv , X 2,e inv . Instead, we want the model to focus on X 1,e inv , X2,einv and not on X e spu. Invariance constraints encourage the model to focus on X 1,e inv , X 2,e inv . In this example, observe that invariant features are partially informative unlike the 2D case (eq. (4)).\nWhy invariance and information bottleneck? We have illustrated through simple examples when the information bottleneck is needed but not invariance and vice-versa. We now provide a simple example where both these constraints are needed at the same time. This example combines the 2D case (eq. (4)) and the example we highlighted in the paragraph above: Y e ← Xeinv ⊕ Ne, X1,espu ← Xeinv ⊕W e, and X2,espu ← Y e ⊕ V e. In this case, the invariance constraint does not allow representations that use X2,espu but does not prohibit representations that rely on X 1,e spu . However, information bottleneck constraints on top ensure that representations that only use Xeinv are used. We now describe an objective 8 that combines both these principles:\nmin w,Φ ∑ e∈Etr he ( w ·Φ ) s.t. 1 |Etr| ∑ e∈Etr Re ( w ·Φ ) ≤ rth, w ∈ arg min w̃∈Rk×r Re(w̃ ·Φ),∀e ∈ Etr, (6)\nwhere he in the above is a lower bounded differential entropy defined below and rth is the threshold on the average risk. Typical information bottleneck based optimization in neural networks involves minimization of the entropy of the representation output from a certain hidden layer. For both analytical convenience and also because the above setup is a linear model, we work with the simplest form of bottleneck which directly minimizes the entropy of the output layer. Recall the definition of differential entropy of a random variableX , h(X) = −EX [log dPX ] and dPX is the Radon-Nikodym derivative of PX with respect to Lebesgue measure. Because in general differential entropy has no lower bound, we add a small independent noise term ζ (Kirsch et al., 2020) to the classifier to ensure that the entropy is bounded below. We call the above optimization information bottleneck based invariant risk minimization (IB-IRM). In summary, among all the highly predictive invariant predictors we pick the ones that have the least entropy. If we drop the invariance constraint from the above optimization, we get information bottleneck based empirical risk minimization (IB-ERM). In the above formulation and following result, we assume that Xe are continuous random variables; the results continue to hold for discrete Xe as well (See Appendix for details). Theorem 4. IB-IRM and IB-ERM vs. IRM and ERM\n8Results extend to alternate objective with information bottleneck constraints and average risk as objective.\n• Fully informative invariant features (FIIF). Suppose each e ∈ Eall follows Assumption 2. Assume that the invariant features are strictly separable, bounded, and satisfy support overlap (Assumptions 3,5 and 7 hold). Also, for each e ∈ Etr Zespu ← AZeinv + W e, where A ∈ Ro×m, W e ∈ Ro is continuous, bounded, and zero mean noise. Each solution to IB-IRM (eq. (6), with ` as 0-1 loss, and rth = q), and IB-ERM solves the OOD generalization (eq. (1)) but ERM and IRM (eq.(3)) fail.\n• Partially informative invariant features (PIIF). Suppose each e ∈ Eall follows Assumption 1 and ∃ e ∈ Etr such that E[ eZespu] 6= 0. If |Etr| > 2d and the set Etr lies in a linear general position (a mild condition defined in the Appendix), then each solution to IB-IRM (eq. (6), with ` as square loss, σ2 < r\nth ≤ σ2Y , where σ2Y and σ2 are the variance in the label and noise across Etr) and IRM (eq.(3)) solves OOD generalization (eq. (1)) but IB-ERM and ERM fail.\nSignificance of Theorem 4 and remarks. In the first part (FIIF), IB-ERM and IB-IRM succeed without assuming support overlap for the spurious features, which was crucial for success of ERM and IRM in Theorem 3. This establishes that support overlap of spurious features is not a necessary condition. Observe that when invariant features are fully informative, IB-ERM and IB-IRM succeed, but when invariant features are partially informative IB-IRM and IRM succeed. In real data settings, we do not know if the invariant features are fully or partially informative. Since IB-IRM is the only common winner in both the settings, it would be pragmatic to use it in the absence of domain knowledge about the informativeness of the invariant features. In the paragraph preceding the objective in equation (6), we discussed examples where both the IB and IRM constraints were needed at the same time. In the Appendix, we generalize that example and show that if we change the assumptions in linear classification SEM in Assumption 2 such that the invariant features are partially informative, then we see the joint benefit of IB and IRM constraints. At this point, it is also worth pointing to a result in Rosenfeld et al. (2021), which focused on linear classification SEMs (DAG shown in Figure 2c) with partially informative invariant features. Under the assumption of complete support overlap for spurious and invariant features, authors showed IRM succeeds.",
  "4.1 Proposed approach": "We take the three terms from the optimization in equation (6) and create a weighted combination as∑ e ( Re(Φ)+λ‖∇w,w=1.0Re(w·Φ)‖2+νhe(Φ) ) ≤ ∑ e ( Re(Φ)+λ‖∇w,w=1.0Re(w·Φ)‖2+νh(Φ) ) .\nIn the LHS above, the first term corresponds to the risks across environments, the second term approximates invariance constraint (follows the IRMv1 objective (Arjovsky et al., 2019)), and the third term is the entropy of the classifier in each environment.\ngence of |wspu|√\nw2spu+w 2 inv\n(metric from\nNagarajan et al. (2021)) for average selection bias p = 0.9.\nIn the RHS, h(Φ) is the entropy of Φ unconditional on the environment (the entropy on the left-hand side is entropy conditional on the environment assuming all the environments are equally likely). Optimizing over differential entropy is not easy, and thus we resort to minimizing an upper bound of it (Kirsch et al., 2020). We use the standard result that among all continuous random variables with the same variance, Gaussian has the maximum differential entropy. Since the entropy of Gaussian increases with its variance, we use the variance of Φ instead of the differential entropy (For further details, see the Appendix). Our final objective is given as∑\ne\n( Re(Φ) + λ‖∇w,w=1.0Re(w · Φ)‖2 + γVar(Φ) ) . (7)\nOn the behavior of gradient descent with and without information bottleneck. In the entire discussion so far, we have focused on ensuring that the set of optimal solutions to the desired objective (IB-IRM, IB-ERM, etc.) correspond to the solutions of the OOD generalization problem (eq. (1)). In some simple cases, such as the 2D case (eq. (4)), it can be shown that gradient descent is biased towards selecting the ideal classifier (Soudry et al., 2018; Nagarajan et al., 2021). Even though gradient descent can eventually learn the ideal classifier that only relies on the invariant features, training is frustratingly slow as was shown by Nagarajan et al. (2021). In the next theorem, we characterize the impact of using IB penalty (Var(Φ)) in the 2D example (eq. (4)). We compare the methods in terms of |wspu(t)winv(t) |, which was the metric used in Nagarajan et al. (2021); wspu(t) and winv(t) are the weights for the spurious feature and the invariant feature at time t of training (assuming training happens with continuous time gradient descent).\nTheorem 5. Impact of IB on learning speed. Suppose each e ∈ Etr follows the 2D case from equation (4). Set λ = 0, γ > 0 in equation (7) to get the IB-ERM objective with ` as exponential loss. Continuous-time gradient descent on this IB-ERM objective achieves |wspu(t)winv(t) | ≤ in time less than W0( 1 2γ ) 2(1−p) (W0(·) denotes the principal branch of the Lambert W function), while in the same time the\nratio for ERM |wspu(t)winv(t) | ≥ ln( 1+2p 3−2p )/ln\n( 1 + W0( 1 2γ ) 2(1−p) ) , where p = 1|Etr| ∑ e∈Etr p e .\n|wspu(t)winv(t) | converges to zero for both methods, but it converges much faster for IB-ERM (for p = 0.9, = 0.001, γ = 0.58, the ratio for IB-ERM is |wspu(t)winv(t) | ≤ 0.001 and ratio for ERM is | wspu(t) winv(t)\n| ≥ 0.09). In the above theorem, we analyzed the impact of information bottleneck only. The convergence analysis for both the penalties jointly comes with its own challenges, and we hope to explore this in future work. However, we carried out experiments with gradient descent on all the objectives for the 2D example (eq. (4)). See Figure 3 for the comparisons.",
  "5 Experiments": "Methods, datasets & metrics. We compare our approaches – information bottleneck based ERM (IBERM) and information bottleneck based IRM (IB-IRM) with ERM and IRM. We also compare with an Oracle model trained on data where spurious features are permuted to remove spurious correlations. We use all the datasets in Table 2, Terra Incognita dataset (Beery et al., 2018), and COCO (Ahmed et al., 2021). We follow the same protocol for tuning hyperparameters from Aubin et al. (2021); Arjovsky et al. (2019) for their respective datasets (see the Appendix for more details). As is reported in literature, for Example 2/2S, Example 3/3S we use classification error and for AC-CMNIST, CS-CMNIST, Terra Incognita, and COCO we use accuracy. For Example 1/1S, we use mean square error (MSE). The code for experiments can be found at https://github.com/ahujak/IB-IRM.\nSummary of results. In Table 3, we provide a comparison of methods for different examples in linear unit tests (Aubin et al., 2021) for three and six training environments. In Table 4, we provide a comparison of the methods for different CMNIST datasets, Terra Incognita and COCO dataset. Based on our Theorem 4, we do not expect ERM and IB-ERM to do well on Example 1/1S, Example 3/3S and AC-CMNIST as these datasets fall in the PIIF category, i.e, the invariant features are partially informative. On these examples, we find that IRM and IB-IRM do better than ERM and IB-ERM (for Example 3/3S when there are three environments all methods perform poorly). Based on our Theorem 4, we do not expect IRM and ERM to do well on Example 2/2S, CS-CMNIST, Terra Incognita and COCO dataset,9 as these datasets fall in the FIIF category, i.e., the invariant features are fully informative. On these FIIF examples, we find that IB-ERM always performs well (close to oracle), and in some cases IB-IRM also performs well. Our experiments confirm that IB penalty has a crucial role to play in FIIF settings and IRMv1 penalty has a crucial role to play in PIIF settings (to further this claim, we provide an ablation study in the Appendix). On Example 1/1S, AC-CMNIST, we find that IB-IRM is able to extract the benefit of IRMv1 penalty. On CS-CMNIST and Example 2/2S we find that IB-IRM is able to extract the benefit of IB penalty. In settings such as COCO dataset, where IB-IRM does not perform as well as IB-ERM, better hyperparameter tuning strategies should be able to help IB-IRM adapt and put a higher weight on IB penalty. Overall, we can conclude that IB-ERM improves over ERM (significantly in FIIF and marginally in PIIF settings), and IB-IRM improves over IRM (improves in FIIF settings and retains advantages in PIIF settings).\nRemark. As we move from three to six environments, we observe that MSE in Example 1/1S exhibits a larger variance. This is because of the way data is generated, the new environments that are sampled have labels that have a higher noise level (we follow the same procedure as in Aubin et al. (2021)).",
  "7 Conclusion": "In this work, we revisited the fundamental assumptions for OOD generalization for settings when invariant features capture all the information about the label. We showed how linear classification tasks are different and need much stronger assumptions than linear regression tasks. We provide a sharp characterization of performance of ERM and IRM under different assumptions on support overlap of invariant and spurious features. We showed that support overlap of invariant features is necessary or otherwise OOD generalization is impossible. However, ERM and IRM seem to fail even in the absence of support overlap of spurious features. We prove that a form of the information bottleneck constraint along with invariance goes a long way in overcoming the failures while retaining the existing provable guarantees.",
  "Acknowledgements": "We thank Reyhane Askari Hemmat, Adam Ibrahim, Alexia Jolicoeur-Martineau, Divyat Mahajan, Ryan D’Orazio, Nicolas Loizou, Manuela Girotti, and Charles Guille-Escuret for the feedback. Kartik Ahuja would also like to thank Karthikeyan Shanmugam for discussions pertaining to the related works.\nFunding disclosure\nWe would like to thank Samsung Electronics Co., Ldt. for funding this research. Kartik Ahuja acknowledges the support provided by IVADO postdoctoral fellowship funding program. Yoshua Bengio acknowledges the support from CIFAR and IBM. Ioannis Mitliagkas acknowledges support from an NSERC Discovery grant (RGPIN-2019-06512), a Samsung grant, Canada CIFAR AI chair and MSR collaborative research grant. Irina Rish acknowledges the support from Canada CIFAR AI Chair Program and from the Canada Excellence Research Chairs Program. We thank Compute Canada for providing computational resources.",
  "Reviewer Summary": "Reviewer_1: In this paper, the authors re-check the invariance principle and show that out-of-distribution generalization is much harder for linear classification tasks, requiring much stronger restrictions in the form of support overlap assumptions on the distribution shifts. They propose an approach combining both information bottleneck constraints and the invariance principle to addressing the OOD generalization in both regression and classification tasks.\n\nReviewer_2: This work gives an indepth analysis to the sufficient conditions on the generalization of invariant risk minimization (IRM) in classification setting. Authors show that in order to guarantee good generalization, the support overlap condition of invariant features is required. Further, authors also show and validated that if the support overlap condition does not satisfied for spurious features, the combination of IRM objective and information bottleneck regularization (i.e., minimzing the mutual information\nI\n(\nX\n;\nΦ\n(\nX\n)\n)\n) is a reliable remedy.\n\nReviewer_3: This paper investigates the conditions that are needed to provably generalize OOD in the linear classification tasks. Comparing with the linear regression tasks, linear classification tasks need more strong conditions to guarantee the OOD generalization.\nThe paper first proposes the overlap assumptions on both invariant and spurious feature and analyze the OOD generalization guarantee under these assumptions. The theorem shows that only both two types of features satisfy the support overlap assumption can we expect the IRM to solve the OOD generalization problems. Then the paper study the possibility of solving the problem only with the support overlap assumption on the invariant features. The proposed method uses both information bottleneck and invariance constraints. Theoretical results show that the proposed IB-IRM can successfully solve the problem in both fully and partially informative invariant features situations. The experiment results are consistent with the theoretical results.\n\nReviewer_4: The authors first argue that IRM is insufficient to obtain good test performance (only using invariant features) on OOD data at test time when the OOD support is not in the ID support\nThen they prove for a particular distributional assumption on the spurious features, that (additionally) limiting the differential entropy of the representation can solve this issue and lead to OOD generalizing predictors\nIn some toy experiments, they show that IB-IRM sometimes performs better"
}