---
license: mit
tags:
  - neuroscience
  - diffusion-models
  - fisher-information
---

# Multi-scale information geometry: trained checkpoints

Encoder and conditional diffusion checkpoints for the NeurIPS 2026 paper "A multi-scale information geometry reveals the structure of mutual information in neural populations" (Azeglio, Laquitaine, Ferrari and Chalk). Code and project page: https://multigeom.github.io/

## Contents

```
encoders/<encoder>/<area>/best_model.pt      encoder fitted to the recorded responses
encoders/SHA256SUMS                          checksums of the encoders
diffusion/<encoder>/<area>/final_model.pt    conditional diffusion model, 200 epochs
diffusion/<encoder>/<area>/train_log.json    loss per epoch
diffusion/diffusion.yaml                     diffusion training configuration
diffusion/SHA256SUMS                         checksums of the diffusion models
diffusion/README.md                          diffusion training recipe and file format
```

`<encoder>` is `inception`, `resnet50` or `onoff`, and `<area>` is `V1` or `V4`: 12 checkpoints in total.

## Encoders

Each encoder predicts the responses of the 50 most reliable channels of one area, from the dataset of Papale et al. (2025). The Inception and ResNet-50 encoders are frozen backbones with a per-neuron Gaussian readout. The ON-OFF encoders are a frozen front end with a dense linear readout (409,650 parameters).

Each `best_model.pt` is a dictionary with `model_state_dict`, the configuration, and the test-set results used for the encoder table in the paper:

| Encoder | V1 correlation | V1 noise-corrected | V4 correlation | V4 noise-corrected |
|---|---|---|---|---|
| Inception | 0.75 | 0.81 | 0.71 | 0.79 |
| ResNet-50 | 0.78 | 0.85 | 0.70 | 0.78 |
| ON-OFF | 0.20 | 0.22 | 0.13 | 0.14 |

## Download

```
hf download sazio/multigeom --local-dir runs
```

This places the files at the paths the code reads, for example `runs/diffusion/inception/V1/final_model.pt`.

## Loading

```python
import torch

ckpt = torch.load("runs/encoders/onoff/V1/best_model.pt", map_location="cpu", weights_only=False)
```

The files are pickled dictionaries, so load them only from a source you trust. They were saved with NumPy 2.

Paths recorded in the `provenance` and `config` fields were made relative for this release. The weights and results are unchanged.

## License

MIT