A learning method for extended SpikeProp without redundant spikes — Automatic adjustment of hidden units

摘要

In this article, we discuss a leaning algorithm for extended SpikeProp network, which is a kind of spiking neural networks and encodes information by spike timing. Our research group proposed a learning algorithm for extended SpikeProp without redundant output spikes. The performance of the algorithm depends on the network structure. Here, we propose some algorithms that adjust the number of hidden units during its training. Concretely, they remove redundant units one by one. By some experiments, we select the most effective method. It is a method that removes unactive hidden unit when the error is decreased enough. The rate of success trainings is 95% regardless the number of hidden units. And The number of training cycles is less than half of the previous method.