An Energy-Aware Hybrid Particle Swarm Optimization Algorithm for Spiking Neural Network Mapping

Abstract

Listen

Translate article

Recent approaches to improving the scalability of Spiking Neural Networks (SNNs) have looked to use custom architectures to implement and interconnect the neurons in the hardware. The Networks-on-Chip (NoC) interconnection strategy has been used for the hardware SNNs and has achieved a good performance. However, the mapping between a SNN and the NoC system becomes one of the most urgent challenges. In this paper, an energy-aware hybrid Particle Swarm Optimization (PSO) algorithm for SNN mapping is proposed, which combines the basic PSO and Genetic Algorithm (GA). A Star-Subnet-Based-2D Mesh (2D-SSBM) NoC system is used for the testing. Results show that the proposed hybrid PSO algorithm can avoid the premature convergence to local optimum, and effectively reduce the energy consumption of the hardware NoC systems.