Ultra-Low-Power Edge Gating via Adaptive Spiking Neural Network Routers

This technical report evaluates an adaptive power router built from Spiking Neural Network (SNN) units, utilizing dynamic Leaky Integrate-and-Fire (LIF) neurons that gate edge inference systems based on temporal input events. Tested on continuous sensory streams, the adaptive router achieves a 55.89% energy reduction in FP32 (lowering energy consumption from 169.58 mPJ to 74.81 mPJ) while maintaining 99.68% precision, 97.80% recall, and an ultra-low false wake rate of 0.21%. The architecture exhibits robust stability across quantization formats down to INT8 (55.59% energy savings) and graceful degradation under stochastic Gaussian and Poisson noise stress tests up to $1.50$. The accompanying release package includes raw simulation logs (snn_power_router_log.json) and execution scripts (snn_power_router.py) to ensure full reproducibility.