Hierarchical Unsupervised Efficient Coding for Neural Alignment

This repository provides code to initialize a model using hierarchical unsupervised efficient coding, perform supervised fine-tuning, and evaluate the model on brain alignment using the Natural Scenes Dataset (NSD).

The following environmental variables need to be set

NSD_DATA_PATH : e.g. ".../for_atlas"

NSD_STIMULI_PATH: e.g. '.../nsd_stimuli.hdf5'

BACKPROP_TRAINING_PATH : where all backrpop runs are saved, e.g. '.../bottleneck_training_runs'

ENCODING_EVAL_PATH : where all encoding evaluation results will be saved, e.g. '.../deep_bottleneck_data/encoding_eval'

ENCODING_CACHE_PATH : where cached files should be saved to speed up evaluation, e.g. '.../deep_bottleneck_data/encoding_cache'

Usage

Unsupervised Learning

Train the model using hierarchical unsupervised efficient coding:

python training_alternate_layers.py

Supervised Fine-tuning

Launch supervised training using the provided shell script and configuration files:

bash launch_scat_training.sh

Configuration options:

training_params_pca.json — PCA-based initialization training_params_rand.json — Random initialization