dapello/braintree

Regularizing deep net models with primate IT neural data

What is this?

This is the code used to create all models and plots in the ICLR 2023 paper Aligning Model and Primate Late Stage Visual Representations Improves Model-to-Human Behavioral Alignment and Adversarial Robustness

Currently this repo is being cleaned up and modified to include pretrained models and data for training and validation.

Example model training round to produce an IT-aligned CORnet-S:

python main.py -v --seed 0 --neural_loss logCKA --arch cornet_s --epochs 1200 --save_path test_round -nd sachimajajhongpublic -s All -n All -aei \
    --loss_weights 1 1 1 -mix_rate 1 -causal 1 --val_every 30

Note, you will need to download and situate the sachimajajhong dataset (temporarily located at https://ufile.io/dxuc9ghr) for training and the manymonkeys2 dataset (temporarily located at https://ufile.io/1v28f7vy) for validation.

More training scripts to recreate all model conditions and paper results can be found in the array_final*.sbatch files. You will need to install brainscore for these training scripts to work!