Efficient Biologically-Plausible Training of Spiking Neural Networks with Precise Timing

Abstract

Spiking Neural networks (SNNs) are well suited to implementation on energy-efficient neuromorphic processors and to imitation of biological systems. However, promising backpropagation methods have been developed for SNNs, they tend to be either not biologically plausible or to be computationally complex. In this paper we present two biologically plausible alternatives to backpropagation while retaining high temporal precision for SNN training. We show distinct tradeoffs between complexity, accuracy, and amenability to deployment on a neuromorphic processor of several training methods. Finally, We suggest exploration of biologically plausible methods to enable low complexity training on resource-constrained neuromorphic hardware.