Task-structured Modularity Emerges in Artificial Networks and Aligns with Brain Architecture

Contents

• Overview
• Repo Contents
• System Requirements
• Installation Guide
• Training
• Plot Results
• License
• Issues
• Citation

Overview

This is the official repository for Task-structured Modularity Emerges in Artificial Networks and Aligns with Brain Architecture.
In this study, we demonstrate that multitask and incremental learning enhance modularity in recurrent neural networks (RNNs) compared to single-task learning, revealing how functional demands influence the structural organization of neural networks.

Repo Contents

• Data: Data from The Human Connectome Project.
• Python: Main Python source code (see main.py, models/, utils/, etc.).
• Shell scripts: Shell scripts are used to automate and manage multiple
  parallel Python tasks (see Fig2.a.sh, Fig2.b-h.sh, Fig3.a.sh, etc.).

System Requirements

Hardware Requirements

• A modern CPU or GPU (NVIDIA recommended for deep learning tasks)
• Sufficient disk space for data and model checkpoints

Dependencies

• Python 3.9
• PyTorch 1.13.1
• NumPy 1.23.5
• SciPy 1.13.1
• Bctpy 0.6.1
• (See requirements.txt for full list)

Installation Guide

1. Clone the repository:

   git clone git@github.com:brain-intelligence-lab/modular-brain-rnn.git
   cd modular-brain-rnn

2. Create and activate a virtual environment:

   conda create -n mod_rnn python=3.9
   conda activate mod_rnn

3. Install dependencies:

   pip install -r requirements.txt

4. Optional packages

   These packages are only necessary for some analyses and supplementary experiments mentioned in the paper.

   For Linear Mixed-Effects Analysis (Fig.2h) & Bipartite Network Community Detection:

   conda install rpy2
   R
   # Inside the R environment, install required packages:
   install.packages("lme4")
   install.packages("RLRsim")
   # Type 'quit()' to exit R
   

   For GNN Training, :

    # We need to install torch_geometric 
    # We first install pyg_lib torch_cluster torch_scatter torch_sparse torch_spline_conv 
    # from https://data.pyg.org/whl/torch-1.13.0%2Bcu117.html
   
    # Take pyg_lib for example:
    wget https://data.pyg.org/whl/torch-1.13.0%2Bcu117/pyg_lib-0.4.0%2Bpt113cu117-cp39-cp39-linux_x86_64.whl
    pip install pyg_lib-0.4.0+pt113cu117-cp39-cp39-linux_x86_64.whl 
   
    # Then we can install torch_geometric using:
    pip install torch_geometric

Training

To reproduce the results, the shell scripts are here to automate and manage multiple parallel Python tasks.

To start single task learning, use Fig2.a.sh to train all 20 tasks independently:

./scripts/Fig2.a.sh

To start multi-task learning, use Fig2.b-h.sh to train multiple tasks together:

./scripts/Fig2.b-h.sh

For other experiments in the paper, just use the corresponding scripts.

Plot Results

Once training is complete, you can generate the figures used in the paper by running the corresponding Python plot scripts:

python plot_Fig2.py

Results traning log and figures can be found in the ./runs and ./figures respectively.

NOTE: Before each training session, please clear or rename the previous training's directory (e.g., mv ./runs/Fig2a ./runs/Fig2a_pre) to prevent the SummaryWriter from appending new log data, which could lead to incorrect plotting.

Overall Workflow

Citation

If you use this code or data, please cite:

@article{YourCitation2025,
  title={Task-structured Modularity Emerges in Artificial Networks and Aligns with Brain Architecture},
  author={Your Name and Collaborators},
  journal={Journal Name},
  year={2025}
}