---
library_name: torchvision
pipeline_tag: feature-extraction
license: other
tags:
  - resnet-50
  - robust-vision
  - neural-encoding
  - feature-visualization
  - activation-maximization
---

# Spatial Neural Feature Accentuation checkpoints

Runtime artifacts for
[Animadversio/spatial-neural-feature-accentuation](https://github.com/Animadversio/spatial-neural-feature-accentuation).

## Files

| File | Purpose | SHA-256 |
|---|---|---|
| `resnet50_robust_backbone.pt` | Adversarially robust ImageNet ResNet-50 state dict | `6c6731b622d6e521d4e36707f5a0d24d18ff7d8ffee1ac30b7a68eb36871c763` |
| `resnet50_robust_25_compiled_targets.pt` | 25 PCA/readout objectives collapsed to feature-space weights and biases | `69975cfdf5abaaecfaf34ea76d05b02f93bc2d4f13a195d99b3652a1f7b24757` |

The compiled cache contains five selected neural-encoding targets for each of
five monkeys (`leap`, `paul`, `red`, `three0`, and `venus`). Internal filesystem
paths have been removed from this publication copy. It retains target IDs,
subject labels, unit IDs, robust-ResNet layer names, effective weights/biases,
and q01/q99 response normalization values.

## Use

The companion repository downloads these files at a pinned Hub revision and
checks both SHA-256 hashes before loading them. They can also be downloaded with:

```bash
hf download binxu/spatial-neural-feature-accentuation-checkpoints \
  --include 'resnet50_robust_*.pt' \
  --local-dir checkpoints
```

These files are intended for differentiable feature visualization and the
research-art workflow documented in the companion repository. The 25 compiled
targets are not general-purpose image classifiers.

## Licenses and provenance

The companion code is MIT licensed. These weight artifacts retain the terms of
their original models and source data; users are responsible for complying with
those terms. See the companion repository for method details, limitations,
privacy guidance, target definitions, and reproducibility metadata.