emergent-ei-reservoir-networks

---
name: emergent-ei-reservoir-networks
description: "Emergent Excitatory-Inhibitory structure in performance-evolved reservoir computing networks for neuronal population dynamics. Uses PDNE framework to evolve compact reservoirs from Wilson-Cowan systems, recovering E-I sign patterns without explicit design. Activation: excitatory inhibitory, E-I balance, reservoir computing, Wilson-Cowan, neuronal population dynamics, PDNE, network evolution, digital twin neuron."
---

# Emergent E-I Structure in Performance-Evolved Reservoir Networks

> Performance-driven network evolution (PDNE) applied to Wilson-Cowan neuronal systems produces compact reservoir computing networks that spontaneously recover excitatory-inhibitory sign patterns from data alone, enabling structurally interpretable digital twins of neuronal systems.

## Metadata
- **Source**: arXiv:2603.13635
- **Authors**: Manish Yadav
- **Published**: 2026-03-13
- **Category**: nlin.AO (Adaptation and Self-Organizing Systems)

## Core Methodology

### Key Innovation
The paper demonstrates that performance-driven network evolution can produce **structurally interpretable** models of physiological rhythms. Starting from minimal seed networks, the PDNE framework grows and prunes reservoir computing networks based solely on prediction performance — and the evolved networks spontaneously recover the correct E-I sign pattern of the Wilson-Cowan model for 3 of 4 interaction types, without this being imposed by design.

### Technical Framework

#### 1. Wilson-Cowan (WC) Model
The canonical two-population model of excitatory-inhibitory interaction:
- Two coupled ODEs for excitatory E(t) and inhibitory I(t) population activities
- Underlies physiological rhythms (alpha, gamma oscillations)
- Four interaction types: E→E, E→I, I→E, I→I

#### 2. Performance-Dependent Network Evolution (PDNE)
- **Initialization**: Start from a minimal seed reservoir network
- **Iterative Growth**: Add nodes/connections to improve prediction accuracy
- **Pruning**: Remove low-contribution connections to maintain compactness
- **Fitness**: Prediction performance on WC dynamics (no structural bias imposed)
- **Termination**: When prediction accuracy plateaus

#### 3. Structural Analysis
- Node specialization: E-specialized, I-specialized, and shared E-I representation nodes
- Connectivity sign analysis: Examine whether evolved connections recover WC sign patterns
- Generalization testing: Zero-shot transfer to novel stimulus configurations

### Key Results
1. **Accurate Prediction**: Evolved reservoirs accurately predict both E(t) and I(t) across unseen stimulus amplitudes
2. **Zero-Shot Generalization**: Transfer to novel pulse number, position, and amplitude without retraining
3. **Structural Recovery**: Population-level connectivity recovers correct E-I sign patterns for 3/4 interaction types
4. **Compact Models**: Performance-driven evolution yields minimal sufficient networks

## Implementation Guide

### Prerequisites
- Python with NumPy, SciPy for ODE integration
- Reservoir computing framework (e.g., PyReservoir or custom implementation)
- Wilson-Cowan model implementation

### Step-by-Step

1. **Implement Wilson-Cowan Model**
```python
import numpy as np
from scipy.integrate import odeint

def wilson_cowan(state, t, params):
    E, I = state
    w_ee, w_ei, w_ie, w_ii = params['weights']
    tau_e, tau_i = params['time_constants']
    # Sigmoid activation
    S = lambda x: 1 / (1 + np.exp(-x))
    dE = (-E + S(w_ee * E - w_ei * I + params['stim_e'])) / tau_e
    dI = (-I + S(w_ie * E - w_ii * I + params['stim_i'])) / tau_i
    return [dE, dI]
```

2. **Initialize Seed Reservoir**
```python
# Minimal reservoir with small random connectivity
n_nodes = 10  # Start small
W_res = np.random.randn(n_nodes, n_nodes) * 0.1
W_in = np.random.randn(n_nodes, 2)  # 2 inputs: E, I
```

3. **PDNE Evolution Loop**
```python
for generation in range(max_generations):
    # Evaluate prediction performance on WC dynamics
    performance = evaluate_reservoir(W_res, W_in, wc_data)
    
    # Grow: add nodes if performance below threshold
    if performance < target:
        W_res, W_in = add_nodes(W_res, W_in, n_new=2)
    
    # Prune: remove low-contribution connections
    W_res = prune_weak_connections(W_res, threshold=0.01)
    
    # Analyze emergent structure
    ei_structure = analyze_ei_signs(W_res)
```

4. **Structural Analysis**
```python
def analyze_ei_signs(W_res):
    """Check if evolved connectivity recovers E-I sign patterns."""
    # Classify nodes by their specialization
    # Compare connectivity signs to expected WC pattern
    # Expected: E→E (+), E→I (+), I→E (-), I→I (-)
    pass
```

## Applications
- **Digital Twins of Neuronal Systems**: Compact, data-efficient models that capture both dynamics and structure
- **Neuroscience Model Discovery**: Inferring E-I structure from neural recordings without prior assumptions
- **Brain-Computer Interfaces**: Lightweight models for real-time neural dynamics prediction
- **Neuromorphic Computing**: Evolved compact network architectures for hardware implementation
- **Epilepsy Research**: Understanding E-I balance disruption through structural analysis

## Pitfalls
- PDNE may not recover all interaction types — the I→I connection sign was not consistently recovered
- Wilson-Cowan is a simplification; real neuronal systems have more complex dynamics
- Reservoir size and initialization sensitivity may affect convergence
- Zero-shot generalization has limits — extreme stimulus changes may require adaptation

## Related Skills
- neural-population-dynamics
- reservoir-computing
- wilson-cowan-model
- excitatory-inhibitory-balance
- network-evolution
- snn-working-memory-heterogeneous-delays