Date of Award

1999

Degree Type

Thesis

Degree Name

Bachelor of Engineering Honours

Faculty

Faculty of Communications, Health and Science

First Advisor

Abdesselam Bouzerdoum

Abstract

This project investigates a new class of high-order neural networks called shunting inhibitory artificial neural networks (SIANN's) and their training methods. SIANN's are biologically inspired neural networks whose dynamics are governed by a set of coupled nonlinear differential equations. The interactions among neurons are mediated via a nonlinear mechanism called shunting inhibition, which allows the neurons to operate as adaptive nonlinear filters. The project's main objective is to devise training methods, based on error backpropagation type of algorithms, which would allow SIANNs to be trained to perform feature extraction for classification and nonlinear regression tasks. The training algorithms developed will simplify the task of designing complex, powerful neural networks for applications in pattern recognition, image processing, signal processing, machine vision and control. The five training methods adapted in this project for SIANN's are error-backpropagation based on gradient descent (GD), gradient descent with variable learning rate (GDV), gradient descent with momentum (GDM), gradient descent with direct solution step (GDD) and APOLEX algorithm. SIANN's and these training methods are implemented in MATLAB. Testing on several benchmarks including the parity problems, classification of 2-D patterns, and function approximation shows that SIANN's trained using these methods yield comparable or better performance with multilayer perceptrons (MLP's).

Share

 
COinS