Early Image Termination Technique During STDP Training of Spiking Neural Network

Abstract

Spiking Neural Network (SNN) is a breed of neural networks that seek to achieve low energy and power by more closely mimicking biological brains. SNNs are often trained using lightweight unsupervised learning such as Spike Time Dependent Plasticity (STDP). However, STDP is prone to redundant time steps during training since STDP cannot determine current image needs further training or not. To reduce redundant time steps and lower energy costs during STDP training, we propose a novel technique that terminates training upon an image preemptively. The proposed technique reduces time steps by 44% with accuracy drop of 0.91% on MNIST.