SWAT: An unsupervised SNN training algorithm for classification problems

Abstract

The work presented in this paper merges the Bienenstock-Cooper-Munro (BCM) learning rule with Spike Timing Dependent Plasticity (STDP) to develop a training algor ithm for a Spiking Neural Network (SNN), stimulated using spike trains. The BCM rule is utilised to modulate the height of the plasticity window, associated with STDP. The SNN topology uses a single training neuron in the training phase where all classes are passed to this neuron, and the associated weights are subsequently mapped to the classifying output neurons: the weights are proportionally distributed across the output neurons to reflect similarities in the input data. The training algorithm also includes both exhibitory and inhibitory facilitating dynamic synapses that create a frequency routing capability allowing the information presented to the network to be routed to different hidden layer neurons. A variable neuron threshold level simulates the refractory period. The network is benchmarked against the non-linearly separable IRIS data set problem and results presented in the paper show that the proposed training algorithm exhibits a convergence accuracy comparable to other SNN training algorithms.