Learning Compositional Generalization from Biological Plasticity Dynamics

Central Thesis: Structure is Learning. Artificial Neural Networks typically start dense and learn weights. Biological brains start sparse and learn structure (topology). This project demonstrates that applying biological structural priors (derived from neuroplasticity simulations) to AI models leads to efficient, self-organizing sparsity without sacrificing accuracy.

Overview

This project bridges Computational Neuroscience and Deep Learning. It consists of three components:

1. Biological Simulation: A ground-up simulation of Izhikevich neurons, Spike-Timing-Dependent Plasticity (STDP), and Homeostatic Scaling to model how "engrams" (memories) form in biological tissue.
2. Meta-Learning Optimization: A framework to scientifically derive the "Optimal Neuroplasticity Constant" (connectivity sigma, learning rates) from the biological simulation.
3. Dual-Stack AI Implementation: A validation suite that enforces these derived biological rules in standard Deep Learning models using both PyTorch and TensorFlow.

Key Results

• Derivation: Successfully derived optimal sigma and A_plus parameters that maximize modularity and small-worldness in biological networks.
• Validation: Implemented these rules in a custom PlasticDense layer for TensorFlow and a StructuralPlasticityOptimizer for PyTorch.
• Performance: The bio-regularized AI models achieved ~52% sparsity while maintaining ~98% accuracy on MNIST, proving that the network can self-organize into an efficient topology.

Repository Structure

• simulation/: Pure Python implementation of the biological brain model (Neurons, STDP, Network).
• analysis/: Tools for Topological Data Analysis (Betti numbers, Persistence) to measure network structure.
• metalearning/:
  • layers_tf.py: TensorFlow Custom Layer (PlasticDense) implementing self-contained structural plasticity.
  • model.py / regularizer.py: PyTorch implementation.
• train_benchmark_tf.py: TensorFlow benchmark script.
• train_benchmark.py: PyTorch benchmark script.
• experiment_optimization.py: The meta-learning loop to find optimal biological constants.

Installation

pip install -r requirements.txt

Requires Python 3.8+. Dependencies include tensorflow, torch, numpy, matplotlib, brian2 (optional), ripser (for topology).

Usage

1. Run the Biological Simulation

Explore how plasticity parameters affect network topology.

python experiment_optimization.py

2. Run the AI Benchmark (TensorFlow)

Train a standard MLP vs. a Bio-Regularized MLP to see structural learning in action.

python train_benchmark_tf.py

Check viz/benchmark_result_tf.png for results.

3. Run the AI Benchmark (PyTorch)

python train_benchmark.py

Citation

If you use this code for research in Neuro-AI or Structural Plasticity, please cite:

Suryadevara, R. (2025). Learning Compositional Generalization from Biological Plasticity Dynamics.