Competitive Hyperparameter Balancing on Spiking Neural Network for a Fast, Accurate and Energy-Efficient Inference

Abstract

The Spiking Neural Network (SNN) is currently considered as a next generation neural network model. However, its performance often lags that of classical Artificial Neural Networks. Although there has been a wide range of research to improve the accuracy of SNNs, their performance is determined not only by accuracy, but also by speed and energy efficiency. In this study, we analyzed the relationship between hyperparameters, accuracy, speed and energy of SNN, set a new criterion to estimate the comprehensive performance and applied the Neuroevolutionary algorithm to balance the hyperparameters without the need for manually setting them. The optimized model showed better performance in all terms of our criteria.