Training spiking neurons with gravitational search algorithm for data classification

Abstract

Rather than producing a reaction in its output each iteration, as traditional neurons work, a spiking neuron is excited T ms with an input and actives when a particular value for membrane potential of it obtained. This reaction could possibly be converted to a special firing rate and do a data classification problem based on the firing rate produced by the input signal. Given a set of input instances each belongs to one of the K classes, in this case each input instance is mapped into an input current, then the spiking neuron is excited T ms, and finally the firing rate of input instance is calculated. This model is validated based on next property: data belong to the similar class must produce the same firing rates and data belong to other classes need to produce firing rates adequately different to differentiate among the classes. To provide this property, a training stage id needed to optimize the synaptic weights of model. Gravitational Search Algorithm (GSA) is a novel optimization algorithm designed for solving complex optimization problems. This algorithm has a very much adjusted system for balancing between exploration and exploitation. In this paper, we optimize the synaptic weights of a spiking neuron by GSA. The performance of the proposed algorithm is assessed through four standard benchmark datasets from the UCI Machine Learning Repository. The performance of proposed GSA is compared against the results reported for the same spiking neuron trained with the Differential Evolution (DE) algorithm, the Particle Swarm Optimization (PSO) algorithm, and the Cuckoo Search (CS) algorithm.