competition-stability-functionality-ei-networks

# Competition, Stability, and Functionality in Excitatory-Inhibitory Neural Circuits ## Paper Overview **arXiv**: 2512.05252 (v2, revised 2026-06-03) **Authors**: Betteti, Retnaraj, Davydov, Cortés, Bullo **Category**: Neurons and Cognition (q-bio.NC) ### Abstract Energy-based models rely on symmetric synaptic matrices, excluding biologically realistic excitatory-inhibitory (E-I) networks. This work extends the energetic framework to asymmetric firing rate networks, revealing a **game-theoretic structure** where each neuron acts as an agent minimizing its own energy. Combined with stability principles from network theory, this framework revisits Wilson-Cowan and lateral inhibition models, studying cortical columns as contrast enhancers. ## Core Contributions ### 1. Game-Energetic Interpretation **Key Insight**: Asymmetric E-I networks exhibit competitive dynamics rather than global energy minimization. - Each neuron = rational agent optimizing individual energy - Game theory replaces gradient descent in asymmetric systems - **Nash equilibrium** corresponds to stable network states **Activation Keywords**: `game-theoretic`, `asymmetric networks`, `E-I balance`, `competitive dynamics` ### 2. Network Stability Principles **Rigorous stability analysis** using control theory: - **Small-gain theorem**: Ensures bounded activity in E-I circuits - **Passivity-based stability**: Regulation via inhibitory feedback - **Balance conditions**: E/I ratio constraints for dynamical stability **Mathematical Framework**: ``` - Symmetric networks: Gradient flow → energy minimization - Asymmetric E-I: Game dynamics → competitive equilibrium - Stability: Network control theory (small-gain, passivity) ``` ### 3. Wilson-Cowan Model Extensions **Classical model revisited** with game-theoretic perspective: - Original: Symmetric coupling assumption - Extended: Asymmetric E-I with competitive energy minimization - **Predictions**: - Contrast enhancement in lateral inhibition - Selective sharpening of environmental differences - Hierarchical E/I interplay in cortical columns ### 4. Cortical Column Microcircuits **Functional interpretation**: - Lateral inhibition microcircuits = contrast enhancers - **Hierarchical excitation-inhibition** sharpen subtle differences - **Computational role**: Feature selection via competitive dynamics ## Methodology ### Energy-Based Modeling (Extended) **Traditional approach** (symmetric weights): ``` E(x) = -1/2 x^T W x + Σ f(x_i) ∇E = -W x + f'(x) # Gradient descent ``` **Game-energetic extension** (asymmetric E-I): ``` E_i(x) = -Σ_j W_ij x_j x_i + f(x_i) # Individual energy ∂E_i/∂x_i = -Σ_j W_ij x_j + f'(x_i) # Best response dynamics ``` ### Stability Analysis **Small-gain theorem application**: ``` ||T_E|| · ||T_I|| < 1 # Stability condition where T_E, T_I are excitatory/inhibitory transfer functions ``` **Passivity condition**: ``` ∫_0^t u(s) y(s) ds ≥ 0 # Energy dissipation via inhibition ``` ## Key Findings ### Finding 1: Asymmetric Networks as Games **Claim**: E-I networks are fundamentally game-theoretic systems, not gradient-based optimizers. **Evidence**: - Symmetry relaxation → competitive dynamics - Nash equilibrium = stable firing rate configurations - Biological E/I ratio ≈ 4:1 satisfies stability conditions ### Finding 2: Contrast Enhancement via Competition **Mechanism**: Lateral inhibition sharpening: 1. Excitation: Amplify strong signals 2. Inhibition: Suppress weak neighbors 3. **Competition**: Winner-take-all dynamics 4. **Result**: Enhanced contrast in cortical columns ### Finding 3: Stability from Network Control **Mathematical guarantee**: Small-gain theorem provides bounded activity: - Prevents runaway excitation - Ensures E-I balance maintenance - **Biological plausibility**: Matches observed E/I ratios ## Applications ### Application 1: Neural Circuit Design **Use case**: Engineering stable E-I networks - **Input**: Desired activity pattern - **Method**: Game-energetic optimization - **Output**: Asymmetric weight matrix with stability guarantees ### Application 2: Wilson-Cowan Model Analysis **Use case**: Predicting cortical dynamics - **Input**: E/I connectivity, stimulus - **Method**: Game equilibrium computation - **Output**: Activity trajectory with stability proof ### Application 3: Cortical Column Modeling **Use case**: Contrast enhancement circuits - **Input**: Sensory input distribution - **Method**: Competitive dynamics simulation - **Output**: Sharpened feature representation ## Implementation Patterns ### Pattern 1: Game-Energetic Simulation ```python # Asymmetric E-I network simulation import numpy as np def game_dynamics(W_E, W_I, x_init, T): """ Simulate competitive dynamics in E-I network. W_E: Excitatory weights (asymmetric) W_I: Inhibitory weights (asymmetric) x_init: Initial firing rates """ x = x_init.copy() for t in range(T): # Individual energy minimization (best response) for i in range(len(x)): x[i] = np.maximum(0, W_E[i] @ x - W_I[i] @ x + f_inverse(x[i])) # Stability check (small-gain condition) if not check_stability(W_E, W_I, x): raise ValueError("Unstable configuration") return x # Nash equilibrium ``` ### Pattern 2: Stability Verification ```python def verify_ei_stability(W_E, W_I, alpha=0.1): """ Check small-gain theorem conditions. """ # Excitatory transfer gain T_E = np.linalg.norm(W_E, ord=2) # Inhibitory transfer gain T_I = np.linalg.norm(W_I, ord=2) # Stability condition return T_E * T_I < (1 - alpha) ``` ### Pattern 3: Contrast Enhancement Circuit ```python def lateral_inhibition_contrast(input_signal, W_E, W_I): """ Implement cortical column contrast enhancement. """ # Initial excitation x_excited = W_E @ input_signal # Competitive inhibition x_competitive = game_dynamics(W_E, W_I, x_excited, T=100) # Sharpened output sharpened = x_competitive - np.mean(x_competitive) return sharpened / np.max(np.abs(sharpened)) ``` ## Comparison with Existing Methods | Method | Symmetry | Stability | Biological Plausibility | |--------|----------|-----------|------------------------| | Hopfield | Required | Global minimum | Low (symmetric assumption) | | Wilson-Cowan | Optional | Local analysis | Medium (phenomenological) | | **Game-Energetic** | Not required | Global guarantee | **High (asymmetric E-I)** | ## Limitations 1. **Simplified neuron models**: Firing rate approximation 2. **Static weights**: Plasticity not incorporated 3. **Deterministic dynamics**: Noise effects unexplored 4. **Linear stability**: Nonlinear effects may diverge ## Future Directions 1. **Plasticity integration**: STDP + game dynamics 2. **Noise robustness**: Stochastic game theory 3. **Multi-column networks**: Hierarchical competition 4. **Learning algorithms**: Game-based training for SNNs ## Cross-Domain Connections ### Connection 1: Game Theory → Control Systems - Nash equilibrium = Lyapunov stability - Best response dynamics = Gradient-free optimization - **Reference**: `control/game-theoretic-socio-technical-control` ### Connection 2: E-I Networks → Spiking Neural Networks - Game dynamics → Spike-time competition - Stability → Spike threshold adaptation - **Reference**: `neuroscience/chaos-synchrony-ei-networks` ### Connection 3: Energy Methods → Machine Learning - Game-energetic → Adversarial training - Competitive dynamics → Winner-take-all attention - **Reference**: `ai_collection/winner-take-all-spiking` ## Key References 1. **Wilson-Cowan (1972)**: Original E-I model 2. **Hopfield (1984)**: Symmetric energy-based networks 3. **Bullo et al. (2019)**: Network control theory 4. **Game theory**: Nash equilibrium, best response dynamics ## Activation Triggers Use this skill when: - Modeling **asymmetric E-I neural circuits** - Analyzing **Wilson-Cowan dynamics** with game theory - Designing **stable competitive neural architectures** - Studying **cortical column contrast enhancement** - Investigating **neural circuit stability** via control theory **Keywords**: `excitatory-inhibitory`, `game-theoretic neural dynamics`, `asymmetric networks`, `E-I balance`, `lateral inhibition`, `cortical columns`, `competitive dynamics`, `neural stability`, `energy-based models`, `Wilson-Cowan extension`