Information Processing in Neural Networks

Abstract

As a background to a comprehensive discussion on recent work in neural modelling, the thesis provides a brief biological background description of some of the components of Natural Neural Systems. The review of the work of earlier neural modellers concentrates on the work of the last fifteen years, with the exception of a few earlier, classic papers. The research work performed for this thesis employs an Experimental System evolved to test the ability of a digital, discrete-time model of a neural net to process patterns of signals, provided as input. The aim of the Experiments was to find the type of networks that can perform useful pulse processing functions. The Experimental work is divided into two sections, the first analysing non-adaptive nets and the second analysing networks which use Hebb-type algorithms to alter the strength of interconnections between cells. The first section describes and displays activity of many neural nets. Different inputs are applied to the net and the effects noted. The study is extended by employing Spectral Analysis techniques. The effect of many parameters on frequency of firing of the net are examined including, for example, the decay rates used in the cells and the frequency of the input signal. The second section simulates adaptive nets and examined the relationship between input signals and the final activity of the adapted net. It also employs Spectral Analysis and a specially defined form of display, the Cell Firing Histogram which provides information on how the circuits are being altered by the algorithm. A simple mechanism for recognising signal patterns is proposed that employs several of the properties discovered using the Cell Firing Histogram.