Applying Neural Networks, A Practical Guide
Author: Kevin Swingler
Publisher: Academic Press LimitedREVIEWED BY: William R. Lee
DATE: November 1996
Review
Kevin Swingler's intention in writing Applying Neural Networks was to present a set of techniques that would allow someone in business or industry to apply neural network technology to applications in the real world. It was his goal that a person armed with this book and a neural network simulation software package could run and manage a neural computing based project (V). He also stated an underlying aim was to show that a neural network based project may be planned from start to finish so that model complexity may be manipulated at the points where it is most convenient to do so (15). Through research on the Internet, I discovered Swingler's home page and his announcement that the book was a result of rewriting his thesis for publication.
The book is divided into two parts, I and II. Part I is intended to be a do-it-yourself manual for conducting a neural network project. In part II, Swingler reviews a number of common application areas in which neural networks have been used such as financial prediction, process control and signal processing. Part II is cross referenced to Part I to demonstrate the techniques used (VI).
"Part I - Techniques for building neural networks" contains eight chapters. The first chapter has the nuts and bolts background of neural networks beginning with a definition as follows:
"Neural networks are statistical models of real world systems which are built by tuning a set of parameters. These parameters, known as weights, describe a model which forms a mapping from a set of given values known as inputs to an associated set of values: the outputs (3)."
Swingler relates that traditional artificial intelligence (AI) is based on rules, facts, and inferences. These traditional AI solutions are programmed in LISP and PROLOG languages. In AI fuzzy logic is sometimes used for fuzzy concepts which are derived by design, and it is not a statistical process. Traditional AI which is rule based uses algorithm computing techniques from simple structures like if ... then and repeat ... until(8).
Chapter 1 also states that neural networks are used for predominantly two types of tasks--classification and continuous numeric functions (mapping functions) (4). Neural networks are described as good approximations to a perfect rule or function based approached, but they lack precision and the formality of traditional computing applications. Neural networks are powerful in their own right in providing near perfect approximations to systems that we have insufficient knowledge to program a solution using traditional methods (8). In Applying Neural Networks, Swingler is concerned with one type of neural network known as multi-layer perceptron (MLP) (10). So how does neural computing differ from traditional programming?
| Programming Approach | Neural Computing Approach |
|
|
In the following chapters of Part I, Swingler discusses data encoding, network building, time varying systems, data collection/validation, output and error analysis, network use, and managing a neural network based project. I found that Swingler's company, Neural Innovation Ltd., was also founded as an outcropping of the book. Neural Innovation Ltd. has an educational URL that summarizes the Neural Software Engineering process just discussed. This is probably the heart of the manuscript as Swingler provides structure to the overall process of developing neural solutions.
"PART II -- Review of neural applications" contains only three chapters, but it is only a sampling of the many ways that neural networks may be used for problem solving.
In chapter nine, neural networks and signal processing are shown in the following examples: data preparation, pre-processing for visual processing, neural filters, speech recognition, production quality control, an artistic style classifier, and fingerprint analysis. These sample applications use some variation of signal processing to determine an output.
Financial and business modeling using neural networks is showcased in chapter ten. Examples given were market modeling, price elasticity of demand modeling, data mining, market simulators, financial time series prediction (predicting price changes), forecasting exchange rate prices, and forecasting stock values.
Lastly in chapter eleven, industrial process modeling provides a survey of neural network based solutions. These networks solve process monitoring and control problems such as batch distillation control, power generator control based on demand, gas furnace control, production machine control, microwave oven control, process monitoring, patient monitoring, and predicting driver alertness which was a case study.
Swingler does an outstanding job of covering the definitions of neural networks and their components. His neural network development structure or engineering process is the highlight of the book. Swingler exhibits some upper division mathematics for those interested in the "what are the minute details" aspect. Some areas of the book have a tendency to be scanned because of the technical difficulty. Overall the material presentation is superb. Swingler reached his goal of developing a cook book for design and implementation principles for neural networks.
Having email correspondence with the author on November 1, 1996, at kevin.swingler@psych.stir.ac.uk he emphasized, "the target audience is primarily industrialists and business people who want to improve their operations using this technology." He feels that the book is of value to students who need a solid overview of the techniques they could use if they work in areas developing or implementing neural networks. I find the material presented fits both categories well and could be a relatively excellent teaching book, although the neural network focuses primarily on the multi-layer perceptron network the information is representative of most neural networks.