Pulse pattern training of spiking neural networks using improved genetic algorithm

In this paper, a training method of a spike train for spiking neural networks (SNNs) by the use of genetic algorithms (GAs) is proposed. There have been some reports on training methods of artificial neural networks. However, there are only a few reports on SNNs. SNNs are known to be very similar to biological neural networks because they treat spike trains. SNNs process information based on pulse signals. In SNNs, spiking neurons receive spike pulses from and fire spike pulses to other neurons. The spiking neurons have a characteristic of suddenly changing the membrane potential immediately before and after firing. The characteristics of the potential behavior cause some difficulties in training SNNs. Many currently used training methods of SNNs apply the Hebb rule or gradient methods. However, under the application of the Hebb rule, SNNs sometimes failed the training test. Furthermore, the gradient methods include complicated calculations. In a previous study, we proposed a training method of SNN using a GA. It is reported that a use of not only connecting weights of SNN but also parameters of spike neuron is effective for SNN training. However, the successful rate of the training is not so good. In this paper, we apply an improved GA which adaptively changes the number of offspring and mutation rate according to the diversity of the population to the SNN training. Simulation shows that the improved GA has better training performance than the traditional GA.