A Supervised Method for Learning Precise Timing of Multiple Spikes in Spiking Neural Networks

Abstract

Information in spiking neural network is encoded using the precise timing of spikes as done in biological neural systems. Neurons learn from these spikes based on their timing. It is evident that supervised learning do occur in biological neural systems, but exactly how it occurs and the development of definite mathematical formulation for training these neurons to fire multiple spikes at precise times remains an open research area. Significant strives have been made to formulate supervised spiking neural network learning rules for multi-spiking neurons, however, convergence is not guaranteed in most of these methods when the output spike train contains more than one spike. To this end, a new learning scheme is proposed in this paper which ensures convergence with an increase in the number of spikes in an output spike train. The proposed method elicits a locality concept of spikes and the approximation capabilities of the least squares method to derive a weight update scheme for training multi-spiking neurons using the precise timing of spikes. The performance of the proposed method is evaluated on spike sequence learning and compared with a well-known supervised learning method for multi-spiking neurons, the ReSuMe. The performance is measured using the correlation-based metric and the proposed scheme achieved better accuracy and convergence rates than the well-known learning method for varying learning periods.