Phillip Rowcliffe
Neural Network Perspectives on Cognition and Robotics edited by Anthony Browne, IOP Publishing, £30.00, ISBN 0750304553The turn of a new century has always been a time to take stock of our past, celebrate what we have already achieved, and predict what is to come. It seems only natural that AI scientists would climb aboard this bandwaggon and tout their ideas of what they believe is in store for us in the next millennium.
It was Anthony Browne's hope that by bringing together the leading minds in cognitive science and robotics he could not only highlight the achievements in artificial intelligence to date, but also refresh our hopes at the possibilities that lie ahead; a brave ideal for a science which so many times has failed to deliver on previous promises. Unfortunately, like AI itself, what he has achieved has also fallen short of the mark.
Artificial Intelligence is a fickle science subject to the whims of fashion but one thing that has survived over the years are neural networks and their use in cognitive science. Although so far unable to achieve their initial objectives, neural networks have found their niche, standing strong and unwavering next to whatever the AI community is touting as the current flavour of the month. However, Browne appears to have lost sight of his intended goal, delivering a book which is less of an overview and insight into Neural Nets, and more of a reference manual or, in some parts, an undergraduate text book.
The book is split up into three sections: Representation, Cognitive Modelling and Adaptive Robotics with various experts in those fields each contributing a chapter. The result however, is a disjointed book wandering aimlessly between styles in prose and depth of content. It lacks the cohesive gel that a good editor should bring and is in desperate need of either chapter rewrites or possibly some kind of binding thread between each chapter. In fact, aside from the initial introduction, the editor appears to have abandoned the book leaving the reader with no apparent conclusion, despite his initial premise.
Thankfully the book is not without its good points and the professionalism and focused nature of a handful of the academics gives the book some semblance of structure.
Neural network learning has always been a complex area of study, with ideas about backpropagation, momentum and probabilities enough to confuse the most ardent graduate student. However, Mark Plumbley brings this diverse area of study together by means of a few carefully chosen examples. He explains why and how these new techniques have brought neural networks out of the void it entered during the 1960s and transformed it into a useful method of pattern recognition for data analysis.
Risto Miikkulainen offers a refreshing insight into the use of Neural Networks in Natural Language Processing. His clear definitions of symbolic and subsymbolic representations quickly allow the reader to extrapolate from his introduction to discuss how LISP is used in symbolic representation of objects.
Miikkulainen's passion for his subject is evident not only from his inclusion of Internet links to the source code and demos discussed in the chapter, but with the excitement he unleashes unashamedly in his writing. Unlike some of the other participants in the book, Miikkulainen is not afraid to highlight the limitations of neural networks in his field and goes on to discuss the possibilities and solutions that still lie ahead.
Unfortunately though, one of the pitfalls of trying to do a perspective of a particular topic is that areas which may be of little interest or, in some cases, little relevance, are included as a matter of course. Browne gives equal weighting to these areas and we are subjected to a dry, unimaginative tutorial by Patric van der Smagt on how robots can be taught to perceive their own movements. He presents an incredibly dense mathematical analysis of the subject, and even though the reader is credited with some background mathematical knowledge, it would take a hardened mathematician to not be confused by terms such as Coriolis coefficients and inverse dynamics. The proliferation of these undefined terms and the continuous subdivisions of each of the sections of the chapter (at one point a section was divided into four subsections) make for an unnecessarily difficult and confusing read.
Fortunately however, John Taylor provides us with an exhilarating explanation of the Relational Mind and how experiments are afoot to show that consciousness is a physical property of the brain. He presents a more realistic use of neural networks in helping us to understand exactly what aspects of the neurological make up of the brain are inherent in giving us a sense of self. His clear and informal approach to the subject manages to bring together both the philosophical and engineering issues of this type of experimentation. Hopefully it will serve as an inspiration for all those involved in AI who feel dissatisfied by the apparent lack of expected progress.
Sharkey, Scutt and Damper bring the book to a close by discussing how neural networks are being used in the development of robots. They manage to touch on the fundamental question most people ask when regarding robots: how long before we have a robot as intelligent as ourselves, and all due credit to them, they do attempt to answer it. They present us with a realistic overview of robotics and how the ambitious goal of using neural networks to create an artificial brain might still be possible.
However a few well written chapters are not enough to rescue a book from the depth of obscurity, where this book is no doubt destined. This is a shame since with the advent of genetic algorithms, neural nets need someone to shout from their corner as they still have so much more left to offer.
