Pulsed Neural Networks

Rating:
Summary: relatively current
Review: Pulsed Neural Networks (90's), Artificial Intelligence (80's), Cybernetics (60's and 70's) Telephone Switch Board (10's and 20's) Hydrodynamics (1700 and 1800) it is amazing the names put on cognitive science through the years. This book is a symposium (13 small books) on developing hardware devices capable of replacing or enhancing neurological functions. Using modeling techniques to duplicate biophysical neural pathways can take two forms. The first are math models, which obviously show the relationship between the neurons (virtual reality). The second type is models that build-spiking neurons in real time to which this book is directed. In the first part of the book, a summary of current thought, written by the main compilers of the book (Maas and Bishop) is worth the price alone. The book addresses the question of biological electrical (vs. chemical or genetic) coding, in which the method of information is actually transmitted and received. The compilers have emphasized the chronological event of development with the articles so that the reader does not become lost in which came first. Gravy is given the reader in the form of articles written by researchers in other fields (VLSI) to the point that the reader wonders if one is still reading a book on biophysics. The hard-wired neural net components are then compared to their biological predecessors for the purpose of obtaining usable "dry lab" tools for experiments. ("Dry-wet-electrical lab", "electrical-dry-lab-wet-computer-lab"?). Even though the material contains electrical engineering stuff it is still very readable to biological types and if interested, can muscle through this stuff. The math model development in Matlab is mentioned, but the reference to Matlab's current capabilities in this area is dated (95). Most of symmetries run in the book are older 200 Pentium type machines, and with a faster (650 up) and better busing Matlab's new neural net toolbox can build some interesting stuff (remember however it is still virtual). The "home modeler" can use chap. 7 and 12 as a theoretical basis for stochastic resonance models which the writers, while dealing with stochastic bit-stream overlooked this aspect. However, H.Wilson's Spikes, Decision and Actions is much better. (Matlab interactive). This is a really good book for modelers (reason for the review as opposed to 'me to' reviews). Most of the neural nets and circuits designs are easily modeled in Matlab's Simulink to give real time results similar to those reported. (Whether the results duplicate reality is always a question with these types of models). Flights of fancy (the reason for modeling in the first place, at least the addictive part) can then be implemented according to the capabilities of the reader. The book also discusses "hard wired" CMOS chips available replicating biological systems with plug in units to standard computer I/O units (Motorola, National, and Fuzzytech). However a larger question comes from this book. How can the output of a non-deterministic system be modeled by deterministic model (hardware or otherwise) inputs (H.Wilson)? Without a specific knowledge of the role that neural architecture plays in the phase modulations and oscillatory behavior, how can information be transferred by digital or analog devices duplication neural transmission. As the author puts it in Chap. 12, "Furthermore it is not even clear what the goal of a learning algorithm for pulsed neural nets should be; the goal to learn a function or a function (operator). This book is not a failing because it cannot answer this question. Indeed, the avenues it reviews and discusses opens up many more fields and sparks new uses for the fields it introduces.

Rating:
Summary: good introduction
Review: This book presents a general overview to the growing field of spiking neural networks and their VLSI implementation written by many of the major figures in this field. It begins with several clear explications of the spiking response model which has been recently popularized by Gerstner (who has several good entries in this volume) and why looking at such a model might be a good idea (it really is completely fascinating). As with all neural network theory, understanding the model will require a fairly solid mathematical background. The middle portion of the book is dedicated to VLSI implementation (which I am not involved with so won't comment) and the last chapters present a wide variety of articles from the highly mathematical to advice for digital simulation of such networks. Chapter 10 is by far the most mathematical chapter and presents the analytic results that have been derived for a homogeneous fully connected network. Although this is far from a complete reference it provides a clear explanation of the reasons for this direction and enough good references at the end of each section to get you started. I have continually turned to this book while getting started in my research in this area.