Automated Design of Neural Network Architecture for Classification

Abstract

This Ph.D. thesis deals with finding a good architecture of a neural network classifier. The focus is on methods to improve the performance of existing architectures (i.e. architectures that are initialised by a good academic guess) and automatically building neural networks. An introduction to the Multi-Layer feed-forward neural network is given and the most essential properties for neural networks; there ability to learn from examples is discussion. Topics like traning and generalisation are treated in more explicit. On the basic of this dissuscion methods for finding a good architecture of the network described. This includes methods like; Early stopping, Cross validation, Regularisation, Pruning and various constructions algorithms (methods that successively builds a network). New ideas of combining units with different types of transfer functions like radial basis functions and sigmoid or threshold functions led to the development of a new construction algorithm for classification. The algorithm called "GLOCAL" is fully described. Results from these experiments real life data from a Synthetic Aperture Radar (SAR) are provided.

The thesis was written so people from the industry and graduate students who are interested in neural networks hopeful would find it useful.

Key words: Neural networks, Architectures, Training, Generalisation deductive and construction algorithms.