A learning method for SpikeProp without redundant spikes -automatic adjusting delay of connections

Abstract

SpikeProp, which is proposed by Booij, is a kind of spiking neural networks. It can learn the timing of output spikes, but cannot adjust the number of output spikes. Our research group has discussed the problem and proposed a learning method that can adjust both timing and number of spikes. However, its learning performance depends on the initial network structure (the number of hidden units, delay, the number of sub-connections, and so on). In this article, we discuss the problem, especially the dependency to delay. We proposed the method that removes sub-connections that have unnecessary delay during training. By the proposed method, we successed training more than 87% regardless of the number of initial delays.