SNN Weight Matrix Blocking Algorithm Based on Reinforcement Learning

Abstract

To tackle the high energy consumption of spiking neural networks (SNNs), this work presents a weight matrix block algorithm utilizing reinforcement learning. The method balances energy consumption and compute performance by coupling block, weighting rearrangement and topology rebuilding. The experimental results indicate energy cost drops by approximately 54%, compute efficacy is enhanced and model accuracy maintained. The algorithm takes advantage of a dynamic reinforcement learning system to adaptively adjust the blocking strategy and presents a potential solution for the SNNs' implementation in resource-limited scenarios. The experimental results validate the effectiveness of this approach in balancing energy saving and performance and lay a foundation for low-power neural network applications.