Neural Information Processing and VLSI [Book Review]

In 1989, when I started to work with artificial neural-network hardware, the field was fairly easy to overview: a small stack of research papers contained most of the information. In the past few years, this has changed dramatically: an enormous amount of ideas, theories, chips, and systems have been created and published. Furthermore, work presented at today’s neural network conferences is often related to neighboring fields such as chaos theory, genetic algorithms, cellular automata, fuzzy systems, and artificial life. The breadth of the field, as well as the tremendous amount of published research papers, make it rather difficult for a beginner to get started in the field. The book Neural Information Processing and VLSI does a good job of giving a thorough overview of the field of neural network hardware. Since the book covers a very large field, it cannot go into great depth on each subject. To compensate for that , each chapter has a long list of references which points the reader to more in-depth information. The book is divided into three parts. The first part, entitled “Paradigms and Models,” introduces the reader to neural networks in the form of artificial neural networks and biologically inspired neural networks. The section on artificial neural networks covers back-propagation networks, Hopfield networks, ART networks, Boltzmann machine, time-delay networks, cellular neural networks, and more. These artificial neural networks don’t try to be biologically accurate but make use of many nonbiological ideas from pattern recognition, statistics, and physics. In contrast, the biologically inspired neural networks model biological neural systems (e.g., visual, auditory, motor) to a high degree of accuracy. The section on biologically inspired networks covers various silicon retina designs. Surprisingly, auditory signal processing chips, which constitute a significant part of biologically inspired designs, are not mentioned here. (In the second part, an example from this domain is described.) The authors seem to be particularly fascinated by cellular neural networks, which take up about one third of this first part. The second part , entitled “VLSI D e s ign Techno logy, ” discuss e s the options for implementing neural algorithms in hardware. These options are purely analog, purely digital, and mixed analog/digital. The analog and mixed signal options are illustrated with a variety off building blocks, including multipliers, filter blocks, dynamic and floating gate memories, and winnertake-all circuits. The purely digital option is illustrated with some building blocks, but more so with complete neuroprocessor chips. These digital processor chips include the Adaptive Solutions CNAPS, the Siemens MA16, and the Intel/Nestor Ni1000. The third and largest part , which takes up about half of the book, is entitled “Application and System Prototypin!.” It is this part which distinguishes this book from others. It describes many