Cognitive-Neural-Architecture-Search

Code for the paper Identifying Models With Brain Like Organizational Properties Using Cognitive Neural Architecture Search. See main.py for an example on how to use the evolutionary computation framework.

Requirements

• The code requires the net2brain toolbox to be pre-installed.
• The code uses the NSD dataset through net2Brain, but can be adjusted to use any other fMRI dataset by modifying the paths in evaluation.py and main.py
• by deafult, weights and biases are used to monitor the architecture search process. Please pass your token as an arguemnt for main.py to access the visualizations.

Functionality

This repository contains only the necessary code to reproduce the experiments from the original paper. Including running the neural architecture search using the NSD data-set. The only exception is Effective Robustness calculation, which requires loading the models directly to the effective_robustness.py code in the following repo.

• main.py runs the evolutionary neural architecture search for a specified brain region. Requires specifying evolutionary search parameters such as mutation and cross-over rate, as well as population size.
• config.py defines the search space of the neural networks to be considered (by deafult only Conv and Pool layers of specific kernel sizes).
• run_train_imagenet.slurm creats different seeds that run train_imagenet.py which fintunes the neural networks using the imagenet data-set.
• main.ipynb loads the architectures identified by the CogNAS run and evalutes the brain similarity.

Funding

This project was funded by the HessianAI Connectom grant "Cognitive Neural Systems Architecture Search" (2050120009) led awarded to Dr. Sari Sadiya, Prof. Dr. Visavanathan Ramesh, Prof. Dr. Gemma Roig, and Prof. Dr. Christoph von der Malsburg.