James A. Anderson and Edward Rosenfeld, eds., Talking Nets: An Oral History of Neural Networks. Cambridge, MA: MIT Press, 1998. xi + 434 pp. $39.95 (cloth).
 

Peter Asaro
Doctoral Student
Departments of Philosophy and Computer Science
University of Illinois, Urbana-Champaign

Mailing Address:
Beckman Institute
405 N. Mathews
Urbana, IL 61801
Phone: (217) 352-7737
Email: asaro@uiuc.edu

(1346 words)
 

Talking Nets is a collection of recent interviews with some of the principle participants in neural network research. Opposite page 1, the reader finds a photo of Jerome Lettvin sitting at his desk, bespectacled, smoking, with the ink from his pen bleeding through his shirt pocket. This image offers a wonderful summary of the book: a very human look into a very intense and esoteric field of scientific research. And while there are obvious limitations to constructing a history based solely on oral history, the interviewees in this volume are simply allowed to give their own idiosyncratic, and often fascinating, accounts of their scientific careers. The result is a wonderful set of stories from which the reader must draw their own historic conclusions.

The seventeen interviews include many of the top names in the field: Lettvin, Walter J. Freeman, Bernard Widrow, Leon N. Cooper, Jack D. Cowan, Carver Mead, Teuvo Kohonen, Stephen Grossberg, Gail Carpenter, Michael A. Arbib, James A. Anderson (co-editor of the book), David E. Rumelhart, Robert Hecht-Nielson, Terrence Sejnowski, Paul J. Werbos, Geoffrey E. Hinton, and Bart Kosko. This order follows a seniority by date of birth, ranging from 1920 to1960, which roughly corresponds to the order in which each became involved in the field. Spanning the period from Lettvin's memories of Warren McCulloch and Walter Pitts' collaboration on their 1943 paper which marked the birth of the field, to the stories of several key players in the neural net resurgence of the mid-1980's, the book manages to offer a surprisingly broad view of the field.

The interviews were conducted by the editors in 1993-94, with the exception of Hinton's in 1995-97 and Anderson's auto-biographical non-interview. Each interview begins with questions about the researcher's childhood, and how they got into science. Readers looking for easy generalities about the formation of a scientific psyche will be disappointed at the vast differences in childhood experiences, though some of the individual stories are particularly interesting. One curious commonality is that most of these scientists set out to understand the really deep philosophical questions regarding the nature of the universe and the mind before they finished college. Many even studied philosophy only to be disappointed and turn to mathematics and physics in a search for greater certainty. By graduate school, their personal narratives become immersed in the details of research.

Another commonly recurring theme in these interviews is the issue of priority. In fact, it would seem that almost every major insight since McCulloch and Pitts has been improperly credited to the most recent and popular of several independent discoverers. Most of those interviewed will agree that credit is often given to the wrong people, but it is those who were usurped who appear most concerned to set the record straight.

Apart from asking initially about childhood experiences, and ultimately about where the future of neural nets is heading, each scientist is pretty much allowed to tell their story in their own terms, with little prodding from the interviewers. The result is an idiosyncratic, often awkward, but always personal account of a scientific career. Occasionally a long series of events is recounted without any reference to dates, and the chronology can become obscure. But on the whole, each interview delivers a sense of the distinct personality and experiences of the scientist. The exception to this is the editor's own short auto-biography. While the most thorough and detailed of the accounts, it lacks the casual liveliness found in the proper interviews.

The most compelling of all the interviews is with Paul Werbos, whose story is reason enough to read the book. Werbos tells the hard-luck tale of an academic misfit. After fixing the cost functions in the models the Pentagon used to decide their involvement policies in Vietnam during a summer internship at RAND in 1968, Werbos returned to his graduate studies in economics at Harvard. There, his committee rejected a series of theses, the first of which described the now famous backpropagation algorithm. This left him jobless and destitute, living in a flophouse, unable to afford food, and going door-to-door at MIT seeking a patron for his work, before his committee finally accepted a thesis which used the algorithm to predict conflicts in a geo-political model. He ended up with a job in a political science department, only to have DARPA buy his time from the university and coerce him into refining the Department of Defense's conflict models. Most of the rest of his career involved similar, if less dramatic, struggles with bureaucracies and academic conventions.

In fact, almost all of the scientists in this book report experiencing, at one time or another, some kind of rejection and alienation for pursuing their research. And they all describe themselves as radicals who are now attempting to change or improve the system in ways which would have made their own careers less difficult.

Another unusual interview is with the husband and wife team of Steven Grossberg and Gail Carpenter. In it, they describe their collaboration, and the relationship between their romantic and professional lives together. The result is a rare glimpse of intimacy in science, or at least how intellectual intimacy has resulted in science for this couple.

Talking Nets is not a good introduction to neural networks themselves, however. The discussions quickly become very technical, and in fields ranging from economics to quantum mechanics to mathematics to neurology. Little effort is made to introduce topics or explain concepts. The editors do include a glossary of key terms in the back, but this would be of little help in understanding many of the technical issues that motivated these scientists. The editors also provide many helpful pointers to introductory texts in neural nets, and to the work of each interviewee. Despite this, there is a great deal of interesting material in the interviews on the lives and motives of contemporary scientists which is accessible with no special understanding of the field itself.

As a history, Talking Nets makes no attempt to offer a comprehensive overview of the field, nor draw any historical conclusions. The editors state in their introduction that they wish to leave it to the reader to draw their own conclusions, though Anderson provides his own views on the field in his non-interview. What one is left with is a collection of varied perspectives, along with some orientation on where they are coming from, which capture some vivid glimpses of a complicated and multi-disciplinary field of study. While most of the scientists agree about what the major discoveries have been, they all have slightly different views of the implications of those discoveries. Because the multiple views end up traversing the same ground from different directions, it is possible to see just how different the individual experiences of a scientific history can be.

A well-known challenge facing oral reports in historical methodology is accuracy. The interviews contain at least three different accounts of how a young Pitts was introduced to McCulloch. They all contain Bertrand Russell and Rudolf Carnap and a precocious Pitts, but each describes a slightly different set of events. This is, of course, a great example of the approximate nature of memory-a topic which so many of these scientists have studied. Similarly, it seems everyone has a different impression of Minsky and Papert's (1969) Perceptrons book, and the "scandal" surrounding it.

But one should not read an oral history looking for the facts. Oral history is about capturing the spirit of a scientific movement, and the stories which make up its lore. And in these respects, Talking Nets does an excellent job. This book should be high on the reading list for anyone interested in the history of neural networks, computational neuroscience, mathematical biology, or even artificial intelligence. There are also more than enough connections to work on the biological side of neurons to make this book worthwhile reading for those interested in the recent history of brain sciences. And for young researchers in any field, each interview gives an inspirational tale of what it takes to struggle to the top of an intellectual discipline.