Pioneering brain workers

There are two factors that make scientific biographies worth reading: the anatomy of a discovery and the ability of scientists to weave this into the story of their lives. Without the discovery, the lives of scientists are pretty much like the lives of anyone else. In Talking Nets James Anderson, a neural net scientist, and journalist Edward Rosenfeld have clearly understood these requirements as they press their 17 victims to highlight their contributions in the area of neural networks. The trouble with this subject is that, unlike the DNA story of Watson and Crick, the subject contains many, sometimes competing techniques and methods of analysis. There are no really major discoveries, just the process of useful additions by many people, of which a sample appears in the book. So this book will be most enjoyed by the large number of seasoned workers in neural networks, who will see it as an interesting record of the background and attitudes of some colleagues with whom they share platforms at conferences.

Sadly, the claim that one of the scientific highlights of the book concerns the understanding of the brain does not come off. Most of those interviewed hint at the fact that they were driven by a curiosity about the brain but virtually none claims to have gained much insight, just more respect. Typical is the honest comment by Geoffrey Hinton: "I am disappointed  that we still haven't got a clue what learning algorithms the brain uses". I find that this is a great pity as advances not covered in the book are indeed being made in programmes where neural networks are part of a larger interdisciplinary picture (as, indeed in Hinton's new environment at University College London). What comes through the interviews is that pressures to deliver products, the pressure to get funds, the lure of mathematical analysis, the desire to get credit for innovation, have distracted many of the chosen contributors from following their curiosity about the brain.

For me, it is the biographies of the older interviewees which hold the greater fascination as they get closer to the real dawning of new ideas. For example, Jerry Lettvin of MIT, 74 when he was interviewed in 1994, recalls the roots of the subject in the cybernetics of the early 1940s. There was real excitement then at the idea that engineering analysis could be applied to the nervous system. Lettvin makes my earlier point in one of his closing remarks: "I think that much of it (current work) is beside the point if you want to relate mental activity to nervous activity." He goes on to speak of "delicious problems" but says they require a spirit of adventure rather than mathematical skill. Some contributors read this as an anti-theory point of view, which it is not. Lettvin and many others want to keep a place for imaginative leaps among the pressures to get funding and achieving analytic publications.

One of the younger interviewees, Bart Kosko, represents the "theory first" point of view. He is known for his grasp of the theoretical combination of neural nets and fuzzy logic. He claims to tell anyone starting in the field to "I learn as much math as you can I keep the ass in the seat." Interesting would be his reaction to the warning by another contributor, Carver Mead, against some models of the visual system that miss their target by attempting to be too rigorous: Mead believes that the visual system is so complicated as to require much imagination before analysis.

Contributors are largely American and somewhat unconcerned with the fact that in the 1970s much momentum was maintained outside the United States. As a result of the negative influence of Marvin Minsky and Seymour Papert (of MIT) there was very little progress (with a few major exceptions) in the US during that time. In fact, the younger contributors in the late 1980s were enthused by the feeling that they were rekindling a flame that was wrongly extinguished earlier.

However, outside of the US such work had been going on through the efforts of the likes of Teuvo Kohonen in Finland, Eduardo Caianiello in Italy, Albert Uttley, John Taylor and David Willshaw in the UK, Shun-Ichi Amari in Japan and many others. Of those surviving only Teuvo Kohonen is interviewed. He speaks modestly of his ingenious design of an associative memory and the impossibility of talking of "anything neural" in the US in the 1970s.

In contrast, I enjoyed reading Hinton's account of the liberating effect of going to California from England in the late 1970s: "England was a much smaller more closed community, where there was a substantive view and there was nonsense." Not much may have changed: the funding councils in the UK still need to be convinced that an interdisciplinary crack at the brain is not nonsense.

There are many ghosts in the book. The names of Warren McCulloch and Walter Pitts (who first suggested an electrical model of the neuron), Frank Rosenblatt (who coined the word perceptron), Denis Gabor (early contributor to associative memory models and optimisation methods), Keith Lashley and Donald Hebb (early neural analysts) appear frequently enough to convince anyone that they were the real grandfathers of this field. Another living ghost is John Hopfield, whose elegant papers relating energy to network dynamics in the early 1980s unleashed a flood of work and is mentioned by most of the contributors. The editors report that they were not able to publish a contribution from him for administrative reasons - a great pity.

Judging the book as a whole, it is a bit of a curate's egg. Good are the parts where the interviewees reveal their personal struggles in making their advances: "It's tragic really... although I would often have an idea first, I usually had it too far ahead of its time," says Steve Grossberg. Bernie Widrow points to a paper which is now a "citation classic" which was seen as uninteresting and was about to be rejected when the editor of the journal gave it the benefit of the doubt. Scientists are not protected from the unfair knocks of professional life.

The poor parts of the book are where attempts are made to exaggerate the achievements of the neural network field. One or two of those interviewed speak as if having made their contributions, the future can only be a mere appendix to their work. Nothing could be further from the truth. When it comes to understanding the enormously complex neural activities of the brain, the field of neuro-computing is key and is still wide open. There is massive scope for a new generation to move on imaginatively from the work lovingly described in the pages of this book.

Igor Aleksander is professor of electrical engineering, Imperial College, London.