Neural Nets Are Sparking Heated Debates Among Their Enthusiasts

A scientist’s dream: a “thinking” computer smart enough to recognize objects by sight, understand human speech and respond in kind, even learn by example. And if that’s not enough, the computer’s pattern of thought would help unlock the mysteries of how the human brain works. That’s the promise of the hot new approach to computing that’s been dubbed neural networking— and scientists are paying attention. Neural network courses and doctoral programs at places like the Marine Biological Laboratory and Caltech have been forced to turn away interested graduate students. Industry is funneling millions of dollars into neural net research (The Scientist, October 17, 1988, page 1). And in July, a study done for the Defense Advanced Research Projects Agency (DARPA) suggested that more than $400 million be poured into the burgeoning field during the next eight years. “There’s no question that this is a promising technology and we’re going to scale up our investment,” says Craig I. Fields, DARPA deputy director for research. If they live up to their potential, neural networks will be able to help defense scientists to find stealth air- craft or quiet submarines. Tanks will steer themselves; weapons will find their own targets. But as the field continues to heat up, the stakes rise—and so too the passions. As is so often the case in science, this glamour field is stirring more than mere excitement; it is stirring debate. On the one hand, a turf war has broken out between two rival societies anxious to control the new field (see story, this page). On the other, the very science of neural network research is up for argument. Key scientists are bickering about what form neural networks should take. Unusual, perhaps, in the normal course of science, this philosophical debate over the proper approach to neural network research is not arcane. Rather, it is very much like the age-old question: Does the end justify the means or does it not? Everyone agrees that what makes neural networks especially attractive is that they come close to mimicking the human brain, in both structure and function. The reason: they are made up of interconnected webs of nervelike elements instead of linear circuits like conventional computers. (Or, in their first practical form, they simulate such webs using software on conventional computers.) The debate begins when scientists try to decide what to do with the basic idea. Neuroengineers, for example, care only about whether they can use neural net theory to make headway in their struggle to solve problems—making computers recognize objects, for example. What they don’t really care about is whether the neural net system works in the same way as the brain does when it’s trying to, say, read a landscape. Cursory Attention Other scientists are more interested in how the brain works, but even these researchers disagree about the right tack. Some, like mathematician Cristof Koch of Caltech, argue that researchers should take a “top-down” approach. They view the nervous system as a problem-solver and devise models accordingly, with cursory attention, if any, to biological realities. “Models provide a conceptual framework on which you hang your facts,” explains AT&T Bell Laboratories’ David W Tank, who has developed a neural network-based model for speech recognition. “You don’t want to do it exactly the way it is done in the nervous system. That, I don’t think, gives you any insight into how it works.” In this approach, neurobiological models need only capture the essence, not the details; they need to focus on the algorithms—or problem-solving procedures, not necessarily on how nerve cells themselves carry out these procedures. But this strategy dismays the neurobiologists who have been drawn to the field because of its promise in shedding light on biological functioning (see story, page 5). The neurobiologists tend to see little value in constructing models unless the models actually mimic biological systems. “My quarrel is that models don’t use the information that we do know,” said Nobel laureate Francis Crick during a panel discussion at a September meeting about the brain sponsored by the journal Nature in Boston. “It’s not a real model if it doesn’t have the right characteristics. You’re not producing models, you’re producing caricatures. If you spend too much time on models that are unrealistic, we’ll never know the answers. They [modelers] should take some notice of the biology. Neurobiologist James Bower of Caltech concurs. “What we have to work with is the structure of the nervous system,” he says. “So from an engineering point of view, why ignore it? From a neurobiological point of view, I don’t think you can ignore it.” In fact, argues Duke neurobiologist John Moore, the neural network systems developed by some of the pioneers in the field aren’t worthy of the name neural network. Caltech physicist John Hopfield, for example, who helped breathe new life into the field, has created an artificial neural net that poaches on the traditional forests of neurobiology by possessing a rudimentary memory. But this doesn’t satisfy neurobiologists like Moore. Since Hopfield’s networks depend on flip-flop engineering elements that bear no resemblance to actual nerves, Moore grouses: “They should call their systems parallel processors and delete the words nerve, neuron, and neural. It would be a much more apt description of what they can do. But they clearly wouldn’t be happy doing that because then they would lose the fad factor.” Moore is miffed by the thought of neural network-wielding computer scientists and “top-down” theoreticians invading the domain of neurobiologists. “I’ve known the bulk of people in neurophysiology around the world for years,” says Moore. “All of the sudden, new people are telling me how the brain works.” But he—and other biologists— are angered even more by what might happen to neuroscience because of the sudden popularity of neural networks. Specifically, they fear that large amounts of money from DARPA, which disburses funds without stringent peer review, could lead neuroscientists to abandon valuable approaches to understanding brain function for the chancier, but potentially better-funded field of neural networks. Since the biologists expect that all the hoopla surrounding DARPA’s study could cause NSF and NIH to pour more of their own money into a neural net approach to neuroscience, they worry about the integrity of their field. “I don’t think we should dump more money into it without being certain what we're getting into,” says Bower. Caltech vision researcher David Van Essen agrees. “I don’t know how it will pan out with new players in the game,” he says. “Growth is needed, but it should not be too fast. If there is too much money, the science will be of lower quality.” The rapid growth and divergent approaches have caused sparks to fly. At the Nature meeting in Boston, Terrence J. Sejnowski of the Salk Institute in Sen Diego raised eyebrows—and tempers—when he argued that models are essential because they suggest novel experiments. Offended neurobiologists responded that they would rather abandon science than be experimentalists who are told by modelers what experiments to do and what the results mean. And at a workshop earlier that month in Woods Hole, Mass., bottom-up proponents were quick to attack top-down projects that did not follow what was known about the nervous system. After MIT scientist Alexander Pertland concluded an impressive talk about how computers interpret shapes, he was challenged by Bower. “Why is this important?” Bower asked. Top-down experts, on the other hand, refused to see any importance in distinguishing between real neurons and computer representations in their models. “We have some amazing arguments,” says Bower. “It’s kind of a creative conflict, though we have to be careful to keep our egos out of it.” Models Needed “Things are a little chaotic,” admits Sejnowski “We’re still arguing about basic strategies.” Patricia Churchland, a philosopher from the University of California, San Diego, who has tried to smooth troubled waters, argues that one must start somewhere—and that the two camps should try to work together. “If you don’t have models, then you have this collection of data that threatens to become clutter,” she says. It’s too soon to tell whether the two sides can resolve their differences and learn from each other. But what both camps can agree on is that neural networks can’t possibly do everything claimed for them—at least, not yet. “With any field that has potential commercial significance, there’s so much hype,” says computer scientist David Tank. “A lot of people have unrealistic expectations about what is going to happen.” Biologist Bower, in fact, views the recent IEEE meeting in San Diego and the International Neural Network Society gathering in Boston (see story, page 4) as being in part marketing efforts by self-promoters who may not be looking out for the best interests of science. Some scientists estimate that less than 15% of activities at the meetings involved the presentation of good scientific data and results. “I do this [kind of research], but I think neural net- works are not the panacea that people think they are,” says Koch. “They’re just a tool.” Elizabeth Pennisi is a freelance writer in Tucson, Ariz.

Neural Nets Are Sparking Heated Debates Among Their Enthusiasts

Interested in reading more?

Become a Member of