Abstract

Spiking neural networks (SNNs) possess biological plausibility and energy efficiency as they communicate using asynchronous and mostly sparse spikes. These features make them an ideal choice for efficient neuromorphic computing. The non-differentiable, discrete binary spike events transmitted in SNNs pose a challenge for applying gradient-based optimization algorithms directly to these networks. Therefore, efficient techniques are necessary to enhance energy efficiency without sacrificing accuracy. In this work, we propose Spike Timing Dependent Gradient (STDG), a fast and lightweight learning scheme that uses temporal locality among spikes to avoid non-differentiable derivatives. Our experiments show that STDG reaches the state-of-the-art accuracy of 99.5% and 98.2% on the Caltech101 face/motorbike and the MNIST datasets, respectively.