Skip to content

Home

Jump to bottom
Mo Shakiba edited this page Dec 20, 2025 · 1 revision

Overview

The core of this repository is the Jupyter notebook neuro-constrained-RNN.ipynb, which replicates the key experiments from the paper. It constructs and trains RNN variants constrained by anatomical and functional data from the MICrONS dataset (session 6, scan 6, field 2). These models are evaluated on three cognitive tasks:

  • One-Choice Inference
  • Go/NoGo
  • Perceptual Decision-Making

The notebook initializes weights using biological priors (e.g., STTC, correlations, or precision matrices), applies spatial embedding and communicability regularization during training, and computes metrics such as accuracy, modularity, small-worldness, assortativity, and entropy. Results are saved in a structured folder for analysis, including raw data, summaries, and standard deviations.

Eleven model variants are tested demonstrating that neuronal constraints lead to improved task accuracy and emergent brain-like network properties.

Dependencies

The notebook requires the following Python libraries (installed via pip in the code):

  • bctpy, tensorflow, scipy, matplotlib, numpy, pandas, seaborn, tqdm, networkx

It also downloads the preprocessed MICrONS data (connectome, STTC/correlation/precision matrices, and neuronal coordinates) from a GitHub-hosted ZIP archive.

Usage

  1. Run the Notebook: Execute neuro-constrained-RNN.ipynb in a Jupyter environment (e.g., Google Colab or local setup). The code handles dependency installation, data download, and unzipping automatically.

  2. Key Parameters: Customize the simulation via cell-defined variables (defaults reflect paper settings):

    • simulations: Number of runs per model variant (default: 20).
    • session, scan, field: MICrONS data identifiers (default: 6, 6, 2).
    • number_of_nodes: Neurons in the network (default: 312).
    • network_grid: Grid for random spatial models (default: (12, 13, 2)).
    • sign_constraint: Enforce positive weights (default: False).
    • use_only_sttc: Use only STTC for bio-weights (default: False).
    • use_precision_matrix: Use precision matrix for bio-weights (default: False).
    • TASK1, TASK2, TASK3: Toggle tasks (default: True for all).
    • noise_level: Input noise stddev (default: 0.05).
    • regu_strength: Regularization lambda (default: 0.3).
    • emd_strength: Earth Mover's Distance regularization (default: 0.1).
    • activation_function: RNN activation (default: 'relu').
    • random_network_initialization: Weight init method (default: 'Orthogonal').

⚠️ Note: number_of_nodes and network_grid must match each other and the actual number of neurons in the chosen session/scan/field. Check MICrONS-Data/INF.md.

  1. Execution Flow:

    • Load the preprocessed MICrONS data.
    • Initialize biological weights (lognormal scaling, spectral radius control).
    • Generate task datasets (numpy/tf.data).
    • Train RNN variants with custom regularizers (e.g., spatial embedding, communicability).
    • Compute and save metrics across epochs and simulations.

  2. Outputs: Results are stored in a folder named based on parameters (e.g., 6_6_2_sttc-corr). This includes:

    • Per-task raw CSV files for metrics (e.g., accuracy, modularity).
    • Summary CSVs with means and standard deviations.
    • A ZIP archive for download.

Clone this wiki locally