The neural network model (NN) comprised of relatively simple computing elements, operating in parallel, offers an attractive and versatile framework for exploring a variety of learning
structures and processes for intelligent systems. Due to the amount of research developed in
the area many types of networks have been defined. The one of interest here is the multi-layer
perceptron as it is one of the simplest and it is considered a powerful representation tool whose
complete potential has not been adequately exploited and whose limitations need yet to be
specified in a formal and coherent framework. This dissertation addresses the theory of generalisation performance and architecture selection for the multi-layer perceptron; a subsidiary
aim is to compare and integrate this model with existing data analysis techniques and exploit
its potential by combining it with certain constructs from computational geometry creating a
reliable, coherent network design process which conforms to the characteristics of a generative
learning algorithm, ie. one including mechanisms for manipulating the connections and/or
units that comprise the architecture in addition to the procedure for updating the weights of
the connections. This means that it is unnecessary to provide an initial network as input to
the complete training process.
After discussing in general terms the motivation for this study, the multi-layer perceptron
model is introduced and reviewed, along with the relevant supervised training algorithm, ie.
backpropagation. More particularly, it is argued that a network developed employing this model
can in general be trained and designed in a much better way by extracting more information
about the domains of interest through the application of certain geometric constructs in a preprocessing stage, specifically by generating the Voronoi Diagram and Delaunav Triangulation
[Okabe et al. 92] of the set of points comprising the training set and once a final architecture which performs appropriately on it has been obtained, Principal Component Analysis
[Jolliffe 86] is applied to the outputs produced by the units in the network's hidden layer to
eliminate the redundant dimensions of this space.