Self-Stabilizing Neuronal Control (SSNC) on TREC-6

Overview

SSNC is a biologically-inspired neural network architecture that self-organizes during training. Unlike standard networks with fixed architectures, SSNC adapts its structure dynamically:

• Self-Pruning: Neurons with low utility are gradually "soft deleted" (gated) and eventually "hard deleted" (removed).
• Neurogenesis: When the network is overloaded, new neurons are created to add capacity.
• Stability: An Adaptive Observer monitors the network to ensure changes are principled and safe.

This project demonstrates SSNC on the TREC-6 Question Classification dataset.

Project Structure

The project is modularized for clarity:

/
├── main.py                 # Entry point to run training
├── train.py                # (Legacy/Alias)
├── data/                   # Dataset storage
├── results/                # Logs and visualization outputs
└── src/
    ├── ssnc_mlp.py         # Core SSNC network implementation
    ├── observer.py         # Adaptive Observer logic (The "Brain")
    ├── metrics.py          # Neuron state tracking
    ├── trec6_model.py      # TREC-6 specific model wrapper
    ├── trec6_loader.py     # Data loading utilities
    ├── trainer.py          # Training loop
    └── utils.py            # Helpers

Quick Start

1. Install Dependencies:

   pip install torch numpy matplotlib tqdm

2. Run Training:

   python main.py

3. View Results: Check the results/trec6/ directory for:

   • trec6_stable_results.png: Plots of accuracy, loss, and neuron counts.
   • network_architecture.png: Visualization of the final network structure.
   • trec6_stable_history.json: Raw training logs.

Documentation

For a deep dive into how SSNC works, read ARCHITECTURE.md.