Tracking The Progress Of Today's Brain 'Technology'

NEUROCOMPUTING: Foundations Of Research James A. Anderson and Edward Rosenfeld, editors MIT Press; Cambridge; 729 pages; $55 Brain science, neural computation and traditional artificial intelligence, perhaps more than most fields, seem to lack definitive textbooks. Instead they give rise to classic papers or multiauthored compendia such as this volume, which fits in squarely with other such important multiauthored landmark works as Principles of Neural Science, Handbook of Artificial Intelligence, 1986’s Parallel Distributed Processing, and the Morgan Kaufmann Readings in ... series. In this volume, Anderson and Rosenfeld have achieved three major and difficult goals brilliantly. First, they have endeavored to bring together in one place most of the “must-read” papers which truly are “foundations of research,” and perhaps more important, of current and future practice. A second major goal in a historical collection such as this is to give some idea of the development and intermingling of different lines of research over time. This too has been quite successfully accomplished, especially for some of the much-cited but little-read work predating the mid-1960s. The arrangement of the papers is roughly chronological, and the reader gets a sense of the progression of neural network research not only from the editors’ excellent introductory comments, but also from the inclusion of both earlier and later papers by the same authors. It is particularly interesting to watch the development of ideas in subsequent papers by McCulloch and Pitts, Kohonen, Grossberg, Anderson, Hopfield, and others as one reads through the volume. It’s occasionally surprising as well; for example, the shift away from the simple threshold logic units introduced by McCulloch and Pitts in the 1943 paper, which everyone cites, to the more topographic neuronal organization of their less well-known 1947 paper, which is also included. A third task for a collection such as this is to present a great deal of material from disparate sources within a unified point of view. The editors are quite explicit about their point of view, supplying a philosophical context firmly grounded in brain science (rather than computer science) to the papers they have selected. For them, it is clear that the brain is not constructed to think abstractly, but instead to ensure survival in the world, and consequently is fitted to the environment. To make the fit as close as possible within the limited resources available, the brain is full of special-purpose hardware, hacks, and kludges, which ruin the aesthetic clarity of some of the neural theory they implement. Anderson and Rosenfeld almost always address the engineering virtues of papers and experiments they have selected to present here, and ruthlessly place them above theoretical elegance. A common axiom of the engineering outlook— “steal when you can”—is markedly evident in their repeated admonitions that neural network practitioners should be thoroughly familiar with brain hardware, draw on the brain for theoretical inspiration, and measure the success of their experiments by how well they predict how the brain actually acts, as demonstrated by neurophysiology, timing experiments, and so forth. In other words, “steal from the brain when you can.” In this light, it is to their credit that they have also included the back-propagation work of Rumelhart, Hinton, and Williams—a great engineering trick that is almost certainly not the way the brain works. Other evidence of an engineering viewpoint is apparent in the thorough coverage of topics like linear associative memories, topographic maps, “thermodynamic” models and annealing schedules, and special-purpose architectures, such as the character- and word-recognition models of Fukushima, McClelland, and Rumelhart. While none of these techniques or architectures is fully general, they will all have a part to play in the development of practical—and profitable—neural network applications in the future. The editors’ underlying conviction that engineering drives forward understanding perhaps as much or more than big science does, will receive ample room for confirmation in the next few years as neural network applications begin to become integrated into practical systems. As neural network consultant Tom Schwartz often puts it, “old technology is never driven out by new technology, it’s just reconfigured.” As has been the case with expert systems technology, I’m sure the “new” technology will itself be “reconfigured” as integration issues come to the fore. What changes will be necessary to adapt a neural network model for practical hybridization with expert systems and distributed databases in the applications of the future? I only hope observers as sharp as Anderson and Rosenfeld will be gath- ering up the results of this next wave for a volume as essential as the one under review here. Joe Bosurgi is a senior member of the advanced technology group at Symantec Corp., a Cupertino, Calif software development firm.

Tracking The Progress Of Today's Brain 'Technology'

Interested in reading more?

Become a Member of