Spike-Aware Training and Timing Window Optimization for Energy-Efficient Inference in Conversion-Based Spiking Neural Networks

Abstract

Spiking Neural Networks (SNNs) are a promising alternative to traditional Deep Neural Networks (DNNs) due to their ability to operate in low-power event-driven mode. However, training SNNs from scratch remains challenging, and conversion-based SNNs derived from pre-trained DNNs have become popular. In this paper, we focus on generating learnable parameters for the inference phase by analyzing the timing window of rate-coded spiking activation using N-MNIST digit classification. We compare the training accuracy of a non-spiking ANN model with the spike ignored and spike-aware spiking activation models trained at different time intervals. We also use regularization to control the mean spike rate of neurons and include a moving-average pooling layer to improve classification accuracy. We provide insights into optimizing the timing window of rate-coded spiking activation for energy-efficient and accurate SNN inference. Our results show that spike-aware training with regularization and moving-average pooling improves convergence and achieves high accuracy. These findings can help improve the training of SNNs for various AI applications.