REVIEW: AN INTRODUCTION TO NEURAL COMPUTING.

G. William Moore, MD, PhD.
http://www.medparse.com/rvneuroc.htm


Send comments and correspondence to: George.Moore4@va.gov
See also: http://www.medparse.com/gwmcv.htm ............. http://www.medparse.com/rvgodell.htm ............. http://www.medparse.com/rvcognis.htm .............
http://www.netautopsy.org/rvflatte.htm

United States Government Work, uncopyrighted, public-domain. This document does not necessarily represent the views or policies of any United States Government agency. This document is provided "as is", without warranty of any kind, express or implied, including but not limited to the warranties of merchantability, fitness for a particular purpose and noninfringement. In no event shall the authors be liable for any claim, damages or other liability, whether in an action of contract, tort or otherwise, arising from, out of or in connection with the document or the use or other dealings made with the document. Published in: Neurocomputing. 2001 Jan;42(1):.



Aleksandr I, Morton H.
An Introduction to Neural Computing. Second Edition.
London: International Thomson Computer Press. 1995.
ISBN 1-85032-167-1, 284 pages.

      This book is an overview of neurocomputing that surveys a complex and growing area of study on a level accessible to the advanced undergraduate in science, engineering, or mathematics. A NEURAL NET, as readers of this journal are well-aware, is a set of connected nodes, or NEURONS, consisting of input-nodes, output-nodes, and internal or 'hidden' nodes. The goal of a neurocomputing system is to train a neural net optimally to predict known outputs for a given set of inputs (training set); then to apply this training information to as-yet-unexamined inputs (test set). The scope, complexity, and relative weighting of the connections, and the application of these models to practical computing problems, form much of the intellectual content of neurocomputing.

      The book begins with the historical origins of neurocomputing: the range and limitations of the Turing-von-Neumann single-processor computer design; the McCulloch-Pitts model of neural nets; and the inadequacy of traditional statistical and numerical analysis models for managing complex data analysis. The book follows the historical development of neural nets through the Minsky-Papert concept of the multilayered perceptron, Hopfield's contributions to associative computer memories, and the theoretical and practical aspects of parallel computing and very large system integration. The book is replete with worked-out mathematical examples and general questions for discussion, including the authors' own responses to the questions, collected at the end of the book.

      The book describes an important current debate in neurocomputing: the so-called hard-learning question, namely, whether a neurocomputing training algorithm should be allowed to use only error measurements from the output layer to adjust its internal parameters (the classical view), or whether other information can influence calculations on the neural net. One alternative: jostle the neural net a little bit with artificial 'heat noise', according to a 'Boltzmann temperature model', akin to annealing in metallurgy. Another alternative: add more nodes, more connections, and adjust the weights among these objects, using some appropriate model. A meta-alternative: design the neural net to recognize general properties of the output. such as edges or branches in a visual output,

      The book clearly makes the case, as it must, against its chief competitor among paradigms for managing large data sets (and possibly for receiving grant funds), namely, artificial intelligence. A neural net is a set of nodes and connections that store experimental knowledge obtained by training on task examples. Artificial intelligence is a set of apriori deductive rulebases. If an artificial intelligence system yields the wrong answer for a particular dataset, then the human software architect must isolate the faulty rule and replace it. Thus artificial intelligence is forever dependent upon human authors and troubleshooters. It is a common experience that an artificial intelligence system dies soon after its creator retires or moves to another institution. The authors missed a chance to comment on a possible borderland between neurocomputing and artificial intelligence: a system grounded in a deductive rulebase but fine-tuned by training on focused task examples. Perhaps either the mathematical complexity or the professional disaffection between the two fields is too great.

      As a physician, I was disappointed by the cursory coverage of neurocomputing in medicine. It's certainly not for lack of funded research or publications over the past several decades. Medical advisory systems, which are briefly described in the final chapter, have been studied for years, and are largely a flop. It's not that computers can't issue reasonable medical advice; but rather that physicians don't need that kind of help from computers. A much more pressing task is to organize and present all the relevant records for a particular patient to the physician; the rest is less of a problem.

      The book concludes with a discussion of Artificial Consciousness. Long the domain of pulp science fiction novels and late-night television shows, this is an area that deserves more attention in the technical literature. Can we settle upon mathematical descriptions of such concepts as self-awareness, representation of meaning, language learning and translation, intention, and emotion? I would add a few of my own: personal privacy, acceptable public disclosure, deceit.

      At the end of the book, the 'Comments on Exercises' was excellent; often I found myself referring to it to help me understand a particular spin or nuance intended by the author. The reference section was adequate, but might have been somewhat larger, considering how easy it is to obtain technical references on the internet. The index was at best fair, which is unpardonable in a modern technical work. I continually found myself remembering that I had read about something earlier in the book, wanting to reread the original section, and being disappointed by the index. Minor flaws, but all in all, a good introduction to neurocomputing.

Last updated: 1/6/2006, by G. William Moore, MD, PhD.