Pulse Neuron Learning Rules for Processing of Dynamical Variables Encoded by Pulse Trains

Abstract

The paper deals with a model of pulse neural network that is applicable for solving of various tasks of processing sensory information. These tasks relate to dynamical variables processing. The distinctive feature of the problem statement is that dynamical variables are represented by pulse (spike) trains. We propose two supervised temporal learning rules for pulse neural network executing the required linear dynamic transformations of variables represented by pulse trains. To generate the required output of the network model we used a reference system with desired properties. The rules minimize the difference between the actual and required pulse train in a local window. The first temporal learning rule was named WB-FILT as it uses the filtered values of errors between binary vectors representing the desired and actual pulse sequences. The second rule was named WB-INST as it uses instantaneous value of the error, which is the difference of the desired and the actual elements of binary vectors. We demonstrated rule’s properties by computer simulation of the mappings of the regular and the dynamical pulse trains. It has been shown that proposed rules are able to configure the simple network that implements a linear dynamic system.