sleep-like-plasticity - SKILL.md Agent Skill

name: sleep-like-plasticity description: Sleep-inspired homeostatic regularization for stabilizing spike-timing-dependent plasticity (STDP) in spiking neural networks. Uses sleep-like phases with modified dynamics to consolidate learning, prevent weight runaway, and improve generalization. Bio-inspired approach mimicking sleep-dependent memory consolidation in biological brains. Use for SNN training stability, continual learning, and bio-inspired regularization. Activation: sleep regularization SNN, homeostatic plasticity, sleep-like consolidation, STDP stabilization, offline consolidation, bio-inspired regularization version: 1.0.0 metadata: hermes: source_paper: "Sleep-Based Homeostatic Regularization for Stabilizing Spike-Timing-Dependent Plasticity (arXiv:2601.08447v1)" published: "2026-01-13" categories: ['cs.NE', 'stat.ML'] authors: Andreas Massey, Aliaksandr Hubin, Stefano Nichele, Solve Sæbø

Sleep-Based Homeostatic Regularization for STDP

Overview

Bio-inspired training technique that introduces sleep-like phases to stabilize STDP learning in spiking neural networks. During sleep phases, network dynamics are modified to consolidate important weights, prune weak connections, and prevent the weight explosion common in pure STDP.

Core Mechanism

Wake Phase (Learning)

Standard STDP learning with input-driven activity
Synaptic weights change based on spike timing correlations
Network encodes task-relevant patterns

Sleep Phase (Consolidation)

No external input; network runs with autonomous dynamics
Homeostatic mechanisms adjust weights toward stability
Synaptic scaling normalizes total synaptic strength
Weak connections are pruned; strong connections are reinforced

Implementation Pattern

class SleepRegularizedSNN:
    def __init__(self, n_neurons, tau_sleep=0.99, weight_floor=0.01):
        self.n_neurons = n_neurons
        self.weights = torch.randn(n_neurons, n_neurons) * 0.1
        self.tau_sleep = tau_sleep
        self.weight_floor = weight_floor
        
    def wake_phase(self, input_spikes, stdp_lr=0.01):
        pre_trace, post_trace = self.compute_traces(input_spikes)
        delta_w = stdp_lr * (pre_trace - post_trace)
        self.weights += delta_w
        
    def sleep_phase(self, n_steps=100):
        for _ in range(n_steps):
            input_norm = self.weights.abs().sum(dim=0, keepdim=True)
            target_norm = torch.ones_like(input_norm) * self.n_neurons
            scale = target_norm / (input_norm + 1e-8)
            self.weights *= scale
            self.weights = self.weights * self.tau_sleep + (1 - self.tau_sleep) * 0.1
            self.weights[self.weights.abs() < self.weight_floor] = 0

Key Benefits

Stability: Prevents weight runaway in STDP
Generalization: Sleep phases reduce overfitting
Energy efficient: Sleep phases require no input data
Biological plausibility: Mirrors sleep-dependent consolidation
Continual learning: Sleep helps integrate new and old memories

References

Andreas Massey, Aliaksandr Hubin, Stefano Nichele, Solve Sæbø, "Sleep-Based Homeostatic Regularization for Stabilizing Spike-Timing-Dependent Plasticity in Recurrent Spiking Neural Networks", arXiv:2601.08447v1, 2026-01-13