Mark Waldron
KBS MSc 13-03-98
Neural networks are for many a fascinating subject. Initially inspired by biological nervous systems they offer the engineer and scientist an excitin= g approach to problem solving and analysis. However the expectation of a solu= tio based on the architecture of the mind is often much greater than the result= =2E This book is representative of a more modern approach to neural networks, namely that of statistical pattern recognition. This approach offers a much more direct and principled route to many of the same concepts. The aim of t= he book is to ground the initiate; be he scientist, student or industrial proj= ect supervisor, in the basics of practicalities of the field. Emphasis is not b= ased on a detailed understanding of the mathematical concepts but rather the des= ign of a system utilising neural network techniques. Tarrasenko has extensive experience in this field, both academic and industrial and hence is well qualified to provide this long needed overview.
The first part of the book is the mathematical groundwork needed for the re= st of the book. This chapter is short but concise. Three types of network architecture are discussed, theses are perceptrons, Kohonen feature maps an= d radial basis function networks. Each is discussed in turn. The structure of each network is described and then the training methods used, strong emphas= is is based on resulting view of the data and the differing abilities of the different networks. The reader is constantly drawn back to the statistical and geometrical underpinnings of each design, this has the effect of keepin= g the readers feet firmly on the ground. Long winded mathematical derivations are consigned to appendices allowing the reader to form a overall general impression of the different networks without loosing the thread of the discussion whilst battling through mathematical notation. A surprising amou= nt of information is covered within a small space, however many relevant point= s that would seriously effect the performance of the networks are discussed. Tarrassenkos ability to cover so much information in such a small amount of space must be a reflection of his experience in these areas. His prose is t= hat of a plain speaking, practical engineer.
The rest of the book (the majority) is concerned with the task of applying = a neural network technique to the problem at hand. He starts with a discussio= n on how to plan a neural computing project. The typical lifecycle of a neura= l computing project is explained. High-level project structures are proposed which detail every step of a well-practised approach. At each point in this cycle he details common pitfalls and design errors. Particular emphasis is placed on time concerns and the importance of constant testing and experimentation.
Oddly the next section is concerned with identifying an application that could be well served by a neural computational approach. I say oddly becaus= e it seems that this chapter is placed out of order in the book. Despite its peculiar positioning this chapter does have much to say. He looks closely a= t the properties of a problem, emphasising those that suggest a neural networ= k approach. This is a very important chapter and is extremely useful to all t= hat consider neural networks as a part of their problem solving toolbox. It pinpoints the problems that one could be exposed to if neural networks are = the wrong choice and thus provides a framework for assessing their feasibility = as a solution to the problem.
By far the largest section of the book is concerned with the design, testin= g and training of the prototype. An overview of the approach to design is provided as a starting point. This overview itemises the key features that need to be considered. These include pre-processing, input/output encoding, selection of type and architecture, training and testing. Tarassenko then proceeds to expand on these points explaining in detail why they are import= ant and the effects choices made in these areas can have on the overall solutio= n. He chooses not to limit his discussion to the resulting network, considerin= g also the effect that design choices have on the entire project. This is a v= ery practical approach to take and, at least in my case, provides a view of neu= ral network design previously unconsidered.
Throughout this entire section Tarassenko constantly refers to actual examp= les drawn from real problems, he explains the problems represented by each exam= ple and a solution to that problem. This approach of example, problem, solution= is much appreciated as it has the advantage of giving the reader experience as well as theoretical abstraction. Unfortunately there is no rigorous explana= tio of the mathematics employed. I think Tarasenko would have done well to prov= ide us with more mathematical rigour in the form of appendices as he had done i= n previous chapters.
Possibly the most useful part of the book is a section on case studies. Two case studies are presented firstly the classification of sleep states from = EEG data and secondly the prediction of diabetes in Prima Indians. All the knowledge that has been presented in the previous sections is drawn upon he= re. If the process of designing a neural network is still abstract even after t= he rest of the book then this chapter will clear away the clouds and fully gro= und one in the practicalities of performing the task. The studies are presented then discussed, the problems are pinpointed and analysed, the networks test= ed and tweaked. The problems are looked at from many directions with multiple possible solutions. The design of the networks is considered by using domai= n specific knowledge and Kohonen nets and clustering algorithms to visualise = the data. Single layer perceptrons, multiple layer perceptrons and radial basis networks are then trained and tested on the data. The cases take the form o= f an ongoing discussion until all the results are collected and final conclus= ions drawn. An incredibly revealing section that illuminates the process of neur= al network design admirably.
This is an excellent and above all useful book that lives up to its aims we=
ll.
In 132 pages Tarassenko provides an excellent overview of the field of neur=
al
networks. It provides guidelines that will help anyone considering neural
networks as a design solution make best use of their power. I perceive it t=
o be
of benefit to newcomers and experienced practitioners alike. The newcomer w=
ill
gain invaluable lessons across the field and(i expect) the experienced
practitioner a fresh look at problems with new eyes. Most people approachin=
g a
neural network project will appreciate this book for its clarity and scope.=
Reviewed by Mark Waldron Student in the department of cognitive sci= ence Sussex University
