Paper: Neuro-Ensemble Distillation: A Cross-Domain Framework for Stabilizing Biocomputing with Cultured Neurons
Author: Stefan Beierle (Independent Researcher)
Zenodo DOI: https://zenodo.org/records/17259126

This repository contains the supplementary material and code for the NED paper.
NED is a thought experiment bridging biological learning substrates (cultured neurons) with classical ensemble learning and policy distillation from machine learning.

The framework addresses the stability challenge of cultured-neuron systems by:

• Training multiple cultures in parallel.
• Averaging and aligning extracted policies.
• Distilling a robust Consensus Policy.
• Replaying it into new cultures or neuromorphic silicon.

• NED_Paper_v1.0.pdf → Full text of the paper.

• ned_pseudocode.py → Conceptual NED pipeline (Stimulate → Record → Extract → Align → Distill → Replay).
• simulation_basic.py → Runs a simplified LIF-based simulation and exports consensus_model.json.
• heatmaps.py → Visualizes policy noise reduction (individual cultures vs. consensus).
• learning_curves.py → Demonstrates acceleration through replay.
• consensus_model.json → Example Consensus Policy IR with metadata.

• heatmaps.png → Policy heatmaps (Bio Models vs. Consensus).
• learning_curves.png → Replay vs. Naïve learning curves.
• mermaid_cycle.md → Mermaid diagram for generational cycle (A→B→C).

1. Clone repo:

   git clone https://github.com/<youruser>/collateral.git
   cd collateral/code

2. Run the basic simulation:

   python simulation_basic.py

   → Exports outputs/consensus_model.json.

3. Generate heatmaps:

   python heatmaps.py

   → Saves outputs/heatmaps.png.

4. Generate learning curves:

   python learning_curves.py

   → Saves outputs/learning_curves.png.

5. (Optional) Inspect pseudocode pipeline:

   python ned_pseudocode.py

{
  "model_name": "ned_consensus_v1",
  "neurons": 32,
  "parameters": {
    "weights": [[0.01, -0.05, ...]],
    "delays": [[3.12, 2.98, ...]]
  },
  "metadata": {
    "ensemble_size": 5,
    "NEDI": 2.97,
    "input_group_indices": [0,1,2,3,4,5,6,7],
    "output_group_indices": [24,25,26,27,28,29,30,31],
    "stimulus_protocol": "pong_task"
  }
}

• Cortical Labs: GitHub Repository
• Hinton, G. et al. (2015). Distilling the Knowledge in a Neural Network.
• Schmidhuber, J. (1992). Learning complex, extended sequences using the principle of history compression.
• Bengio, Y. (2009). Learning Deep Architectures for AI.

This project is a conceptual thought experiment.
It does not use live biological cultures but provides a computational framework and simulation.
Real-world application with cultured neurons requires dedicated laboratory setups (e.g., Cortical Labs CL1).