Quantizing Spiking Neural Networks with Integers

Abstract

Spiking neural networks (SNNs) are a promising approach to developing autonomous agents that continuously adapt to their environment. Developing low-power SNNs that can be implemented on a digital platform is a critical step to the realization of such agents. One of the most important methods of implementing low-power SNNs requires operating at reduced precision. While traditional computer vision has seen a lot of research examining the trade-offs between precision and model performance, such trade-offs are underexamined for contemporary SNNs. This paper studies the trade-offs associated with learning-performance and the quantization of neural dynamics, weights and learning components in SNNs. Our results show that SNNs trained using only integer fixed-point representations can still retain their accuracy while occupying dramatically lower memory footprints and using only energy-efficient fixed-point arithmetic. We show that the memory usage of SNNs trained with reduced precision weights, errors, gradients and neural dynamics can be downsized by 73.78% at the cost of 1.04% test error increase on the DVS gesture data set.