Constructive neural networks with applications to image compression and pattern recognition

The theory of Neural Networks (NNs) has witnessed a striking progress in the past fifteen years. The basic issues, such as determining the structure and size of the network, and developing efficient training/learning strategies have been extensively investigated. This thesis is mainly focused on constructive neural networks and their applications to regression, image compression and pattern recognition problems. The contributions of this work are as follows. First, two new strategies are proposed for a constructive One-Hidden-Layer Feedforward NN (OHL-FNN) that grows from a small initial network with a few hidden units to one that has sufficient number of hidden units as required by the underlying mapping problem. The first strategy denoted as error scaling is designed to improve the training efficiency and generalization performance of the OHL-FNN. The second strategy is a pruning criterion that produces a smaller network while not degrading the generalization capability of the network. Second, a novel strategy at the structure level adaptation is proposed for constructing, multi-hidden-layer FNNs. By utilizing the proposed scheme, a FNN is obtained that has sufficient number of hidden layers and hidden units that are required by the complexity of the mapping being considered. Third, a new constructive OHL-FNN at the functional level adaptation is developed. According to this scheme, each hidden unit uses a polynomial as its activation function that is different from those of the other units. This permits the growing network to employ different activation functions so that the network would be able to represent and capture the underlying map more efficiently as compared to the fixed activation function networks. Finally the proposed error scaling and input-side pruning techniques are applied to regression, still and moving image compression, and facial expression recognition problems. The proposed constructive algorithm for creating multilayer FNNs is applied to a range of regression problems. The proposed polynomial OHL-FNN is utilized to solve both regression and classification problems. It has been shown through extensive simulations that all the proposed techniques and networks produce very promising result