A training method for SpikeProp without redundant spikes — Removing unnecessary sub-connections during training

SpikeProp, which is proposed by Bohte and extended by Booij, is a type of multi-layer networks of spiking neurons. Our research group has proposed a training algorithm for SpikeProp without redundant output spikes. However, its performance is sensitive to the initial network structure, such as the number of hidden units, the number of sub-connections, and their delays. In this article, we discuss the problem, especially the number of sub-connections. And we propose a method to remove unnecessary sub-connections during training. By simple experiments, we show that the method brings less dependency of training performance on the number of initial sub-connections.