Spiking neural network reinforcement learning method based on temporal coding and STDP

Abstract

A method to train a spiking network to solve a classification task using Spike-Timing-Dependent Plasticity is proposed. Learning is based on the ability of STDP to memorize repeating spike patterns: earliest spikes of pattern contribute to output spike and those weights stay high, while others weights falls down to zero. The output neurons are provided with information on the classes by stimulating the neuron corresponding to the desired class to fire an early spike. The network is single-layer, with competition introduced by inhibitory interconnections. The network consists of leaky integrate-and-fire neurons tuned to provide one spike per pattern. Input data is encoded by Gaussian receptive fields, where earliest spikes contains the most information. The learning method is tested on Fisher’s Iris and Wisconsin Breast Cancer datasets, and results compared to Support Vector Machines, Random Forest and formal neural networks with Adam optimizer.