AugMapping: Accurate and Efficient Inference with Deep Double-Threshold Spiking Neural Networks

Abstract

Spiking neural networks (SNNs) are regarded as one of the promising candidates to overcome the high energy costs of artificial neural networks (ANNs), but the accuracy gap between them is still large on practical tasks. A straightforward yet effective conversion scheme was developed recently to narrow this gap by mapping a trained ANN to an SNN. However, current conversion methods require a relatively large number of time steps and spikes, alleviating the advantages of spike-based computation. In this paper, we propose a new augmented spiking neuron model composed of a double-threshold firing scheme, and it is advanced with the ability to process and elicit augmented spikes whose strength is used to carry the number of typical allor-nothing spikes firing at one time step. Based on this model, a new conversion method called AugMapping is developed. We examine the performance of our methods with both MNIST and CIFAR10 datasets. Our results highlight that the as-proposed methods, as benchmarked to other baselines, are advantageous to accurate and efficient computation with SNNs. Therefore, our work contributes to improving the performance of spike-based computation, which would be of great merit to neuromorphic computing.