Putting Brainpower in a Box

By Christine Gorman

The more time scientists spend designing computers, the more they marvel at the human brain. Tasks that stump the most advanced supercomputer -- recognizing a face, reading a handwritten note -- are child's play for the 3-lb. organ. Most important, unlike any conventional computer, the brain can learn from its mistakes. Researchers have tried for years to program computers to mimic the brain's abilities, but without success. Now a growing number of designers believe they have the answer: if a computer is to function more like a person and less like an overgrown calculator, it must be built more like a brain, which distributes information across a vast interconnected web of nerve cells, or neurons.

Although barely off the production lines, the resulting computers, called neural networks, have already started altering the way people think about artificial intelligence. Researchers at Ford Motor, for example, are exploring the possibility of using neural-network computers to find and fix potential problems in new autos. The U.S. Government last year invested $40 million in neural-network research, according to market analysts. Eventually, proponents say, the new technology will lead to computers that can reprogram themselves to deal with any contingency, in situations ranging from directing combat to planning a sumptuous meal.

Last week 1,600 researchers and artificial-intelligence enthusiasts from around the world gathered in San Diego for the second international conference on neural networks. For five days, they studied the recent advances in this form of artificial intelligence and pondered its bright future. Before long, Tom Schwartz, an industry consultant in Mountain View, Calif., told the crowd, "these machines will be recognized as the steam engines of the 21st century."

Conventional computers function by following a chainlike sequence of detailed instructions. Although very fast, their processors can perform only one task at a time. This lockstep approach works best in solving problems that can be broken down into simpler logical pieces. The processors in a neural- network computer, by contrast, form a grid, much like the nerve cells in the brain. Since these artificial neurons are interconnected, they can share information and perform tasks simultaneously. This two-dimensional approach works best at recognizing patterns.

Instead of programming a neural-network computer to make decisions, its maker trains it to recognize the patterns in any solution to a problem by repeatedly feeding examples to the machine. The computer responds to each example by randomly activating its circuits in a particular configuration. The trainer electronically reinforces any connections that produce a correct answer and weakens those that produce an incorrect one. After as many as several thousand trials, the computer activates only those circuits that produce the right answer. "It works just like a kid," says Farrokh Khatibi, senior product manager at AI Ware Inc., a 3 1/2-year-old neurocomputing company in Cleveland. "It learns and learns, and when you tell it what it did wrong, you hope, it won't do it again."

Neural networks come in all shapes and sizes. Until now, most existed as software simulations because redesigning computer chips took a lot of time and money. By experimenting with different approaches through software rather than hardware, scientists have been able to avoid costly mistakes. At last week's convention in San Diego, several firms introduced the real thing: chips that are actually wired to mimic the nerves in the brain.

Perhaps the toughest competition that neural networks will face in the field of artificial intelligence comes from the so-called expert systems used in medicine, banking, navigation and other fields. Instead of looking for patterns, computerized expert systems distill the decision-making process used by human experts into rules of thumb. Neurocomputer researchers argue that neural networks will eventually prove superior, however, because they can adjust more easily to changes in the nature of the problem to be solved.

Others predict that a combination of neural networks and expert systems could solve problems too tough for either to tackle alone. Since natural intelligence consists of several ways of reasoning, the argument goes, computer engineers will have to design artificial intelligence with more than one way of processing information. Says Esther Dyson, editor of Release 1.0, a computer-industry newsletter: "A neural network will tell the difference between a Russian tank and an American tank, and an expert system will tell whether to shoot."

Despite recent advances, neurocomputing attracts skeptics. Thomas Poggio, head of the Center for Biological Information Processing at M.I.T., insists that proponents of neural networks have exaggerated their computers' smarts. "The only thing they have in common with the human brain is the word neural," he argues. At best, neurocomputers consist of only a few thousand connections -- a very small number compared with the trillions of connections between billions of neurons found in the human brain. "Before trying to duplicate the human brain," Poggio says, "scientists will have to learn far more about the brain than they already know."

With reporting by Scott Brown/San Diego and Thomas McCarroll/New York