Code accompanying our paper on topographic deep neural networks.

🔗 Read the preprint!

• Eshed Margalit 📧
• Hyodong Lee
• Dawn Finzi
• James J. DiCarlo
• Kalanit Grill-Spector*
• Daniel L. K. Yamins*

*co-senior authors

This repository has the following components:

• configs/:
  • analysis_configs/: YAML files that establish the link between checkpoint paths and model names, used during analyses and figure generation
  • config/: Training configuration files for use with the VISSL training framework
• notebooks/: Jupyter notebooks (saved as Markdown files) that reproduce all figures and supplementary figures in the paper
• scripts/: standalone scripts to be run from the command line. Contains a dedicated README explaining each script's purpose.
• spacetorch/: installable Python package providing classes and methods for model training, evaluation, plotting, and figure creation
• train.py: core model training script, modeled after this example in the VISSL repository.

You can download model weights here, under tdann_data/tdann/checkpoints.

Take a look at the instructions for help using the existing models or training your own.

There are a number of datasets used in this work:

The easiest way to get the code to run is to clone the folder structure on the machine the code was tested on. Download the tdann_data folder and place it in a location of your filesystem with plenty of free storage.

Next, set the environment variable ST_BASE_FS to point to that directory, either temporarily:

export ST_BASE_FS=/path/to/downloaded/dir

or permanently, by adding those lines to your shell config file (e.g., .bashrc/.zshrc).

The spacetorch package will look in that folder first to find checkpoints, unit positions, etc. If you have large datasets (like ImageNet) elsewhere on your filesystem, either edit paths.py as needed, or pass different arguments to the Dataset constructors.

To install dependencies, follow the instructions in INSTALL.md.

Please use the following citation for the paper or the code in this repository:

Margalit, E., Lee, H., Finzi, D., Dicarlo, J.J., Grill-Spector, K., and Yamins, D.L.K. (2023). A Unifying Principle for the Functional Organization of Visual Cortex. bioRxiv doi: 10.1101/2023.05.18.541361