| name | discrete-signaling-chaotic-regularization-rnn |
| description | Discrete signaling mediates chaotic regularization in recurrent neural networks - theoretical framework linking microscopic chaos to macroscopic geometry of neural representations. Activation: chaotic regularization, discrete signaling, RNN chaos, neural representation manifold, power-law spectrum, cortical chaos. |
Context
Paper: arXiv:2606.04426 (June 2026)
Authors: Jan Bauer, Christian Keup, Jonathan Kadmon, Moritz Helias
Key Question: How can chaotic neural networks sustain stable, smooth population codes?
Cortical circuits operate in intrinsic chaos where tiny input changes lead to divergent responses. Yet population codes vary smoothly with stimuli. This paper explains the paradox through a theoretical framework linking network dynamics to representation geometry.
Core Methodology
1. Theoretical Framework: Chaos → Representation Geometry
Kernel Methods + Dynamical Mean-Field Theory:
- Chaotic dynamics induce local roughness (sharp distortions at small scales)
- Preserve global smoothness across larger stimulus variations
- This structure acts as intrinsic regularizer: enhances generalization while maintaining expressivity
Key Insight: Chaos creates a multi-scale structure:
- Microscopic: rough, sensitive to perturbations
- Macroscopic: smooth, stable manifolds
2. Power-Law Spectral Signatures
Chaotic networks naturally produce power-law spectral signatures:
- Match experimental observations in cortical recordings
- Provides theoretical basis for observed neural activity patterns
- Links dynamics to recorded data
3. Discrete Signaling Mechanism
Discrete signaling (event-driven communication) mediates the regularization:
- Reduces continuous-time chaos sensitivity
- Enables stable information propagation
- Preserves representational structure
Implementation Steps
Step 1: Model Chaotic RNN
N = 1000
g = 1.5
J = g * np.random.randn(N, N) / np.sqrt(N)
spike_threshold = 0.5
signal_events = []
Step 2: Compute Representation Geometry
stimuli = np.linspace(-1, 1, 100)
representations = []
for s in stimuli:
x = simulate_rnn(J, stimulus=s, dt=0.01, T=10.0)
representations.append(x)
local_roughness = compute_derivative_variance(representations, scale='small')
global_smoothness = compute_manifold_geometry(representations, scale='large')
Step 3: Power Spectrum Analysis
activity = simulate_rnn(J, T=1000.0)
freqs, spectrum = compute_power_spectrum(activity)
alpha = fit_power_law(freqs, spectrum)
Step 4: Discrete Signaling Implementation
def discrete_update(x, J, threshold):
events = []
for i in range(len(x)):
if x[i] > threshold:
events.append((i, x[i]))
return events
Pitfalls
- Chaos Threshold: g > 1 required for chaos; g < 1 is stable regime. Check connectivity strength.
- Time Scale Mismatch: Local roughness appears at fast time scales; global smoothness at slow scales. Analyze separately.
- Kernel Method Limitations: Kernel approximation may fail for very high-dimensional representations. Use dimensionality reduction first.
- Spectral Analysis Artifacts: Finite simulation time causes spectral leakage. Use long simulations (T > 1000 time units).
Verification
- Chaos Check: Lyapunov exponent > 0 indicates chaos
- Power-Law Fit: Spectrum follows f^(-alpha) with alpha ~ 1-2
- Representation Smoothness: Global manifold curvature < threshold, local variance > threshold
- Discrete Event Rate: Spike/event rate matches cortical observations (~10-100 Hz)
Applications
- Explain cortical stability: Why chaotic spiking networks maintain smooth codes
- Design neuromorphic hardware: Use discrete signaling for stable computation
- Analyze neural data: Power-law spectra as signature of chaotic regularization
- Regularization theory: Chaos as intrinsic regularizer, not problem to solve
Activation Keywords
chaotic regularization, discrete signaling, RNN chaos, neural representation manifold, power-law spectrum, cortical dynamics, mean-field theory, kernel method, neural manifold geometry, event-driven network
Related Skills
chaos-freezing-without-plasticity: Alternative chaos stabilization method
spiking-oscillation-mapping: Oscillatory states in chaotic networks
neural-manifold-dynamics-learning: Manifold learning from neural data