Spiking Neural Network Architecture Comparison by Solving the Non-linear XOR Problem

This work presents an analog model for a cellular membrane of a neuron as well as its mathematical analysis. The simple model of a spiking neuron from Izhikevich is used to design two neural architectures. From the electrical analysis the synaptic current is obtained and used as an input parameter to the Izhikevich's model. The first architecture has two neurons on the input layer and one on the output layer, using this architecture it is shown that the neuron models based on the behavior of biologic neurons can be used in a similar manner to the well-known second generation neural networks to solve classification problems. Moreover, in the present work it is shown the superiority of the spiking neuron models, which leads to the second architecture to be conformed by a single neuron. Both architectures are compared performance wise, through the classic problem of separating a non-linear dataset, XOR. The results show that these architectures can be used to the classification or clustering of patterns of features.