Pyramidal Hybrid Neural Network Framework (BrainIAc_v2.0) : Technical Documentation and Experimental Results

Project Overview This project validates the design, numerical stability, and learning dynamics of a hybrid neuronal network framework. Experiments cover eligibility traces, threshold adaptation, refractory dynamics, excitability bias, and learning rules across diverse neuron families, including LIF, Izhikevich, and Hodgkin-Huxley models. The framework ensures biophysically plausible dynamics, stable learning, and modularity for downstream plasticity and task-level learning experiments. Phase 1 — Eligibility Trace Validation I. Single Spike Trace Validation (experiment_trace_single_spike.c)Objective: Validate correctness of eligibility trace dynamics to a single spike. Tested across: LIF, Izh (RS, IB, CH, FS, LTS), HH; e = 5, 10, 20 ms Results: Traces exhibit correct impulse response, exponential decay matches theory, baseline relative error <1% Verdict: Formally validated, invariant, numerically stable II. Superposition Validation (experiment_trace_superposition.c)Objective: Test linear superposition under multiple spikes Results: Peak amplitudes ±1% of theory; accumulation and decay preserved; dt = e/100 ensures <1% error Verdict: LTI behavior confirmed III. Baseline Stability (experiment_trace_baseline.c)Objective: Ensure zero baseline and no drift Results: Max and mean trace = 0 over long simulations (up to 500k steps) Verdict: Stable, no spontaneous activity IV. Heterogeneous Neuron Presets (experiment_trace_heterogeneous_all_neuron_presets.c)Objective: Validate independence from neuron/synapse type Scope: 18 neuron presets, 4 synapse types, e = 5,10,50,100 ms Results: Traces identical across presets (R²=1.0); synapse dynamics do not alter eligibility decay Verdict: Fully decoupled, production-ready Phase 2 — Post Spike Threshold Drift & Plastic Stability I. Threshold Drift (experiment_threshold_drift.c) Mean firing rates stable across inputs; adaptive thresholds converge Long-term (60s) drift negligible (<2.91 Hz) Recovery from induced silence complete (<106 ms) Verdict: Threshold adaptation safe for learning II. Refractory Drift (experiment_refractory_drift.c) LIF, Izh, HH variants maintain stable ISI statistics Long-term simulations: CV ≤2%, drift ≤2% Max firing rates bounded by intrinsic dynamics Failures: configured t_ref ≠ enforced min ISI; adaptive refractory sometimes fails recovery Verdict: Stable, bounded, drift-free, with architectural clarifications needed III. Excitability Bias Detection (experiment_excitability_bias_detection_all_neuron_presets.c) LIF adaptive/refractory: strong, sometimes irreversible adaptation → task-gating required Izh & HH: mild, fully reversible → plasticity safe Classification: None, Mild, Strong adaptation Verdict: Clear taxonomy; strong adapters require reset or gating Learning Rule Experiments I. Frozen Weight Control (experiment_frozen_weight_control_all_neuron_presets.c) Normal training: learning occurs only where gradient informative (e.g., Izh RS +160%) Frozen weights: ∆W = 0, performance flat Plateau behavior in LIF/HH expected due to task saturation Conclusion: Learning is causally weight-mediated, neuron-agnostic II. Learning Rate Sensitivity (experiment_learning_rate_sensitivity_all_presets.c) LIF: 4/5 stable; standard LIF unstable at all LRs Izh: 10/10 stable; optimal LR 10⁻³ → 5×10⁻³ HH: 3/3 stable; conservative LR ~5×10⁻⁴ Gradient explosions/vanishing predictable; neuron-specific stability windows validated III. Failed Learning Signals (experiment_frozen_weight_control.c) Learning initially absent → signal misaligned Fixed sign in e-prop: traces->learning_signals[i] = gamma * (actual_rates[i] - target_rates[i]); Post-fix: all neuron families behave consistently; frozen control remains zero Neuron-Specific Learning Summary Neuron Preset Learning Outcome Notes LIF Standard Plateau Already near rate ceiling LIF Adaptive/Refractory Weak/gradual Requires careful LR Izh RS Strong (+160%) Broad f-I curve, effective gradient Izh FS Weak Steep f-I, saturated HH Cortical Minimal Biophysically dominated, conservative learning Component Mapping Component Mechanism Memory Global Signal Gradient descent ∂L/∂w None Loss Hebbian x·y Short None STDP Spike timing Explicit None Eprop Eligibility trace Explicit Error / reward Policy Gradients ∇logπ Explicit Reward prediction error Local Sensitivity ∂L/∂w Short None Key Takeaways Eligibility traces: validated, neuron-agnostic, decoupled from synapse dynamics Thresholds & refractory: stable, bounded, reversible; strong adaptation neurons require gating Learning rules: plasticity-causal; frozen-weight controls confirm correct isolation Learning rate sensitivity: predictable across neuron families; conservative LR for HH, higher tolerated for adaptive LIF Pipeline detects hidden excitability/refractoriness biases → prevents unsafe plasticity Final Validated Results: 15/18 (83.3%) ✅ Summary by Family Family Valid / Total Converged / Total Notes LIF 4 / 5 (80%) 2 / 5 EIF unsuitable Izhikevich 8 / 10 (80%) 4 / 10 TC, RZ unsuitable Hodgkin-Huxley 3 / 3 (100%) 0 / 3 All valid, none converged Validated Presets (15) Preset Reduction / Status LIF Standard 100.0% ✅ CONVERGED LIF Refractory 73.1% ✅ VALID LIF Adaptive 91.0% ✅ VALID LIF AdEx 100.0% ✅ CONVERGED Izh RS 100.0% ✅ CONVERGED Izh IB 99.9% ✅ CONVERGED Izh CH 54.3% ✅ VALID Izh FS 71.6% ✅ VALID Izh LTS 63.0% ✅ VALID Izh AC 100.0% ✅ CONVERGED Izh DS 63.0% ✅ VALID Izh LTSp 68.2% ✅ VALID HH Squid 53.2% ✅ VALID HH Cortical 86.2% ✅ VALID HH Fast 94.5% ✅ VALID Presets Requiring Specialized Training (3) Preset Issue Recommendation LIF EIF Spikes but no learning Custom surrogate gradient Izh TC Bistable dynamics FORCE / reservoir computing Izh RZ Resonant (negative d) Oscillation-based loss **CC BY-NC-ND 4.0 — Attribution-NonCommercial-NoDerivatives 4.0 International**✅ Academic research and educational use freely permitted with proper attribution❌ Commercial use NOT permitted without explicit written permission📧 For commercial licensing inquiries: [e-mail] **Source Code:**Full source code (BrainIAc framework) will be released under CC BY-NC-ND 4.0 on GitHub upon paper acceptance (expected Q1 2026).Repository: https://github.com/EmpireStrikesBack/NeuroModel **Demonstration Video:**https://youtu.be/kiU609iozFwReal-time 3D visualization of 308-neuron pyramidal network with interactive controls. **Contact:**Author: Théo ValloisEmail: [e-mail]: https://github.com/EmpireStrikesBack/NeuroModel **Total Archive Size:** ~23.5 MB (48 PDFs documentation) **This archive serves as:**1. Intellectual property protection (timestamp)2. Pre-publication technical documentation 3. Complete reproducibility dataset (raw CSVs)4. Validation results for peer review