Learning Method of Recurrent Spiking Neural Networks by Using Particle Swarm Optimization

Abstract

In this paper, we propose a learning method of recurrent spiking neural networks by using particle swarm optimization (PSO). The existing learning method by using the gradient based method sometimes falls into local minima. The proposed method aims to find the global optimum regardless of initial solutions. Since PSO can treat optimization problems in which the objective function is a non-differentiable function, in this paper we formulate learning problems with such objective functions and propose a learning method based on PSO to solve them. The proposed method makes it possible to treat not only the learning problem in terms of firing instants which the conventional method treats, but also one in terms of the number and frequency of firings.