Towards Efficient and Stable Time Parameter Optimization in Spiking Neural Networks

In recent years, Spiking Neural Networks (SNNs) have attracted widespread attention in research due to their energy-saving characteristics and ease of deployment. However, the inference performance of SNNs is closely related to the number of time steps, which leads to a trade-off between accuracy and efficiency. Currently, there is still a lack of research on optimizing the time parameters in SNNs. Therefore, this study is dedicated to achieving a balance between time efficiency and performance optimization in SNNs. To this end, we propose an innovative Momentum Performance Queue (MPQ) strategy for optimizing time parameters. MPQ dynamically adjusts the time parameter T during training based on historical performance to maintain or even improve network performance while reducing computational resource consumption. We have validated this method on multiple standard datasets and achieved good results. For example, we achieved an accuracy of 96.07% on the CIFAR-10 dataset using only 1 time step.