Abstract
There is an increasing interest in emulating Spiking Neural Networks (SNNs)on neuromorphic computing devices due to their low energy consumption. Recentadvances have allowed training SNNs to a point where they start to compete withtraditional Artificial Neural Networks (ANNs) in terms of accuracy, while atthe same time being energy efficient when run on neuromorphic hardware.However, the process of training SNNs is still based on dense tensor operationsoriginally developed for ANNs which do not leverage the spatiotemporally sparsenature of SNNs. We present here the first sparse SNN backpropagation algorithmwhich achieves the same or better accuracy as current state of the art methodswhile being significantly faster and more memory efficient. We show theeffectiveness of our method on real datasets of varying complexity(Fashion-MNIST, Neuromophic-MNIST and Spiking Heidelberg Digits) achieving aspeedup in the backward pass of up to 150x, and 85% more memory efficient,without losing accuracy.