Delayed Spiking Neural Network and Exponential Time Dependent Plasticity Algorithm

TL;DR — A realistic delayed spiking neural network is introduced in this study, and a biologically plausible exponential time-dependent plasticity algorithm is proposed to train it.

Abstract

Spiking Neural Networks (SNNs) become more similar to artificial neural networks (ANNs) to solve complex machine learning tasks. However, such similarity does not bring superior performances but loses biological plausibility. Moreover, most learning methods of SNNs follow the pattern of gradient descent used in ANNs, which also suffer from low bio-plausibility. To address these issues, a realistic delayed spiking neural network (DSNN) is introduced in this study, which only considers the dendrite and axon delays as the learnable parameters. And a more biologically plausible exponential time-dependent plasticity (ETDP) algorithm is proposed to train the DSNN. The ETDP adjusts the delays according to the global and local time differences between presynaptic and postsynaptic spikes, and the forward and backward propagation time of signals. These biological indicators can surrogate the time-consuming computation of descents precisely. Experimental results demonstrate that the DSNN trained by ETDP achieves very competitive results on various benchmark datasets, compared with other SNNs.