Pyramidal Hybrid Neural Network Framework (BrainIAc_v2.1) : Technical Documentation and Experimental Results

Project Overview This project validates the design correctness, numerical stability, and learning scalability of a hybrid spiking neuronal network framework implementing e-prop–style local learning rules. The work systematically separates solver stability, biophysical dynamics, and learnability under local credit assignment. Experiments cover eligibility traces, threshold adaptation, refractory dynamics, excitability bias, and learning behavior across diverse neuron families, including Leaky Integrate-and-Fire (LIF), Izhikevich, and conductance-based Hodgkin–Huxley (HH) models. The framework is designed to be modular, biophysically grounded, numerically stable, and suitable for downstream plasticity and task-level learning experiments. Phase 1 — Eligibility Trace Validation I. Single Spike Trace Validation (experiment_trace_single_spike.c) ObjectiveValidate correctness of eligibility trace dynamics in response to a single spike. ScopeLIF, Izhikevich (RS, IB, CH, FS, LTS), Hodgkin–HuxleyEligibility time constants: 5, 10, 20 ms Results Correct impulse response for all neuron families Exponential decay matches analytical solution Baseline-relative error < 1% VerdictFormally validated, invariant, numerically stable II. Superposition Validation (experiment_trace_superposition.c) ObjectiveTest linear superposition under multiple spike events. Results Peak amplitudes within ±1% of theoretical values Accumulation and decay preserved dt = e/100 ensures <1% numerical error VerdictLinear time-invariant (LTI) behavior confirmed III. Baseline Stability (experiment_trace_baseline.c) ObjectiveEnsure absence of drift or spontaneous activation. Results Mean and max trace values remain exactly zero Long simulations up to 500k steps VerdictStable, drift-free baseline IV. Heterogeneous Neuron Presets (experiment_trace_heterogeneous_all_neuron_presets.c) ObjectiveVerify independence of eligibility traces from neuron and synapse type. Scope 18 neuron presets 4 synapse models Eligibility constants: 5, 10, 50, 100 ms Results Traces identical across all presets (R² = 1.0) Synapse dynamics do not affect eligibility decay VerdictFully decoupled; production-ready Phase 2 — Threshold, Refractory, and Excitability Stability I. Threshold Drift (experiment_threshold_drift.c) Results Mean firing rates remain stable across input regimes Adaptive thresholds converge reliably Long-term drift over 60 s < 2.91 Hz Recovery from induced silence < 106 ms VerdictThreshold adaptation is safe for learning II. Refractory Drift (experiment_refractory_drift.c) Results LIF, Izhikevich, and HH variants maintain stable ISI statistics Long-term CV ≤ 2%, drift ≤ 2% Maximum firing rates bounded by intrinsic dynamics Notes Clarified distinction between configured refractory time and enforced minimum ISI Adaptive refractory requires recovery safeguards VerdictStable, bounded, drift-free with documented architectural constraints III. Excitability Bias Detection (experiment_excitability_bias_detection_all_neuron_presets.c) Results LIF adaptive/refractory models show strong, sometimes irreversible adaptation Izhikevich and HH models exhibit mild, fully reversible adaptation Classification None / Mild / Strong excitability bias VerdictClear taxonomy established; strong adapters require gating or reset during training Learning Rule Experiments I. Frozen Weight Control (experiment_frozen_weight_control_all_neuron_presets.c) ObjectiveConfirm that learning is causally mediated by synaptic plasticity. Results Normal training: learning occurs only where gradient is informative Frozen weights: ΔW = 0; performance flat Plateaus in LIF/HH consistent with task saturation ConclusionLearning is weight-mediated and neuron-agnostic II. Learning Rate Sensitivity (experiment_learning_rate_sensitivity_all_presets.c) Results LIF: stable across most presets; broad learning-rate tolerance Izhikevich: universally stable; optimal LR ≈ 10⁻³–5×10⁻³ HH: stable only under conservative LR ≈ 5×10⁻⁴ ConclusionNeuron-specific stability windows exist but are predictable and bounded III. Failed Learning Signal Diagnosis (experiment_frozen_weight_control.c) IssueLearning initially absent due to sign error in e-prop signal. Fix learning_signals[i] = gamma * (actual_rates[i] - target_rates[i]); Post-fix Results All neuron families behave consistently Frozen-weight controls remain zero ConclusionPipeline reliably detects learning-signal misalignment Phase 6 — Multi-Neuron Scaling Experiment (Key Result) QuestionDoes learning scale predictably with network size across neuron families? Scaling Results by Family LIF Family (5 presets) — Fully Successful Presets Standard, Refractory, Adaptive, EIF, AdEx Observed 100% loss reduction for all LIF variants except AdEx at very small N EIF and adaptive variants converge cleanly Negligible inter-trial variance AdEx reaches full convergence by N = 10 Conclusion E-prop is fully compatible with LIF dynamics, including exponential and adaptive variants. Scaling is stable, predictable, and parameter-robust. Izhikevich Family (8 presets) — Fully Successful Observed RS, IB, AC: near-perfect convergence (~100%) CH, FS, LTS, DS, LTSp: consistent 50–90% reduction Low variance; no catastrophic failures Conclusion Reduced convergence in complex subtypes reflects optimization difficulty, not instability. Hodgkin–Huxley Family (3 presets) — Systematic Scaling Failure Observed Squid: 9–28% loss reduction; high variance Cortical: failure at N = 1; partial learning at larger N (<50%) Fast: mixed behavior; inconsistent across N Interpretation Numerical stability already validated Failure is systematic and reproducible Indicates mismatch between local credit assignment and stiff conductance-based dynamics Final Scaling Summary Family Presets Scale Successfully LIF 5 100% Izhikevich 8 100% Hodgkin–Huxley 3 0% Total 16 81% Scientific Conclusions Eligibility traces are correct, neuron-agnostic, and decoupled from synapse dynamics Threshold and refractory mechanisms are stable, bounded, and reversible Learning is causally plasticity-driven, confirmed by frozen-weight controls E-prop scales robustly across all LIF and Izhikevich neuron families without tuning Hodgkin–Huxley failures reflect learning-rule limitations, not numerical instability These results cleanly separate numerical correctness from learnability, establishing a rigorous validation pipeline. Outlook Conductance-based HH neurons likely require: Modified eligibility traces Adaptive learning rates Hybrid local–global credit signals This work identifies learning-rule extensions—not solver fixes—as the correct direction for future research. Archive Purpose This Zenodo archive serves as: Intellectual property timestamp Pre-publication technical documentation Complete reproducibility dataset (raw CSVs, plots, logs) Validation material for peer review LicenseCC BY-NC-ND 4.0 — Attribution-NonCommercial-NoDerivatives 4.0 InternationalAcademic and educational use permitted with attributionCommercial use requires explicit written permission ContactAuthor: Théo ValloisEmail: [e-mail]: https://github.com/EmpireStrikesBack/NeuroModel Source CodeFull BrainIAc framework to be released upon paper acceptance (expected Q1 2026) Demonstration Videohttps://youtu.be/kiU609iozFw Archive Size~24.5 MB (50 PDFs documentation)