In an earlier review (review 9301-0009), I compared introductory books on artificial neural networks (ANNs) published during the last couple of years. A common weakness of all nine books included in this earlier review was the lack of accompanying ANN simulator software. This review focuses primarily on recently published books that do come with ANN software. Seven of these books are sold with the software (mostly for the PC), and the features of the software packages are summarized in Table 1.

Some earlier ANN books, while they did not include software, went to the trouble of providing code listings for simple neural network models. One of the best features of Pao’s book [1], for example, was its appendices on the generalized delta rule and clustering algorithm (C-code listings).

Eberhart and Dobbins

This book enables readers to develop their own simulator software; while it does not boast accompanying ANN software, it does provide extensive code listings. After a historical introduction to ANNs, the authors describe the backpropagation and self-organization models. The following chapters address systems considerations, software tools (with emphasis on C for a DOS environment), development environments (including a description of the authors’ own CaseNet), hardware implementations (focusing on the transputer), performance metrics, network analysis, and expert systems (mainly fuzzy systems). The last five chapters describe the following ANN case studies: detection of EEG spikes, radar signal processing, futures market forecasting, optical character recognition, and music composition.

As is often the case with collections of papers, these chapters vary in quality, the overall effect being rather patchy and inconsistent. The most relevant sections are the four appendices, which provide C-code (shareware) listings for a generic backpropagation network, Kohonen’s self-organizing network, an OCR_Shell that relies on pcx_tp routines from the Genus toolkit, and a music generator program. Only the first two codes are suitable for general use and for porting to different platforms; both are clearly written and well commented.

Müller and Reinhart

The accompanying ANN simulator software is the best feature of this book. It is divided into three parts: “Models of Neural Networks,” “Statistical Physics of Neural Networks,” and “Computer Codes.” While it is mathematically rigorous, it is easy to read and explains the underlying ANN principles well. Throughout the text, a PC icon identifies those portions of text that are supported by the accompanying software. Part 3 not only provides a theoretical background of the relevant ANN models but also serves as a user manual for the simulator software. The applications used to illustrate the workings of the four main ANN models are clear and instructive. All in all, this work is a good mixture of textbook and ANN software simulator.

Korn

Writers of ANN software simulators are a more appropriate audience for this book than readers who wish to supplement their reading on ANNs with computer demonstrations. Desire/NeuNet is an interpretive matrix language adapted for developing ANN software. (“Desire” stands for “direct executing simulation in real time.”) After a brief introduction to ANN theory in chapter 1, Korn discusses the seven matrix language keywords sufficient for building various ANN models. The remaining five chapters cover perceptrons and backpropagation, including XOR (exclusive-OR), parity, and function learning; competitive learning, including character recognition and function learning; more advanced neuron modeling; associative memories; and adaptive predictor, combining ANNs with dynamic systems.

A substantial learning curve is required to make the best use of Desire/NeuNet. A thorough reading of the text is necessary prior to using the software. In addition,  Desire/NeuNet  insists on specific path names and directories under DOS. The best demonstration is the Kohonen net, which provides a good illustration of dynamic network growth.

Blum

A better book for those programmers wishing to develop their own ANN simulator software is Blum’s. After introducing the fundamental concepts of object-oriented programming (class, object, information hiding, inheritance, dynamic binding, and so on), Blum discusses the fundamentals of ANNs. Chapter 3 discusses specific ANN models (backpropagation, counterpropagation, bidirectional associative memory, and Hopfield). The fourth chapter discusses typical ANN applications (stock market prediction, handwritten digit recognition, a spelling checker, and the traveling salesperson problem). The book concludes with two appendices, the first describing the ANN class library, and the second providing a listing of the four applications mentioned earlier. A diskette containing source code from the two appendices is also available from the publisher, but at an additional cost (roughly as much as the book itself).

Kosko

Unfortunately, most ANN simulators include rather idiosyncratic and non-obvious user interfaces, much in the manner of McClelland and Rumelhart’s books [2,3]. A notable exception in this regard is Kosko’s book. The software included on the two 5.25-inch diskettes comes from three different sources: Togai’s target tracker, which compares fuzzy and Kalman filtering approaches; Hyperlogic’s fuzzy differential competitive learning (inverted pendulum) and truck backer-upper; and Olmstead and Watkins’s OWL. OWL has the best user interface of the three, and enables the user to investigate the workings of five ANN models listed. The Hyperlogic inverted pendulum demonstration is particularly good, but both it and the Togai demonstration are more snippets from commercial packages, rather than learning tools per se. In summary, the ANN simulator is reasonable but limited. Where Kosko does fall down is in the text proper--it makes for difficult reading, is heavily endowed with mathematics, and is obviously pitched at a graduate rather than a novice audience. Moreover, it reflects Kosko’s own bias toward fuzzy systems rather than ANNs.

McCord-Nelson and Illingworth

This book is one of the worst I have come across on ANNs; its coverage is glib, superficial, and almost comic book–like. The accompanying software is undoubtedly its best feature, but even it leaves a lot to be desired. The authors dress up their software within an expert system shell--Texas Instruments’ Personal Consultant Plus--arguing that this configuration overcomes some of the nasties of ANN input/output, namely the usual requirement for normalization and/or scaling. The only ANN model included is multilayer perceptron, but the user is not able to learn much about the workings of this ANN; rather, we are limited to consulting a neural expert system to obtain decisions on what model Chevy corresponds most closely to our specifications, or whether a proposed loan is approved.

Orchard and Phillips

The authors come from a cognitive psychology or computational neuroscience background, so it is not surprising to find them referring to McClelland and Rumelhart [2,3]; terms such as “parallel distributed processing” are used throughout their book. While it is short, it provides a good introduction to ANNs. The text also makes reference throughout to those sections that can be further demonstrated using the accompanying software simulator. Furthermore, chapter 2 serves as a well-written user guide for this software. The user interface is good and the software is straightforward.

Durbin, Miall, and Mitchison

Like Eberhart and Dobbins, this book is a collection of papers. Its orientation is biological, and in this sense it is more aligned with Orchard and Phillips and with McClelland and Rumelhart). The book’s 21 chapters are each written by a different author, and as such it has a patchy finish overall. As with most of the other books in this review, its best feature is the accompanying ANN simulator software. Strangely enough, this software bears no particular relation to the book. Nevertheless, it is good enough to stand on its own merits. It is the only ANN simulator besides that of Orchard and Phillips that runs on the Macintosh rather than the PC. It incorporates a good user interface and is easy to use. It provides a total of 11 different ANN demonstrations, each of which has its own accompanying MacWrite READ_ME text file. The pole balancer and topological map demos make particularly good use of the high-resolution graphics available on the Macintosh. The book is mediocre, but the software is commendable.

Comparison

If one is looking for a good introductory book and software package, the only one I recommend is Müller and Reinhart. On the other hand, if you have more of a programming bent and want to write your own ANN simulator, you would probably be best advised to purchase Blum’s book.