IMAGINE a computer that can think. It would learn by trial and error. As it gained experience with something, it would draw increasingly sophisticated conclusions. Computer scientists are turning that science-fiction idea into actual machines.
They're called neural networks. University researchers and computer companies are finding lots of ways to use them. For example:
Speech recognition. At Rensselaer Polytechnic Institute in Troy, N.Y., researchers play hundreds of hours of tapes for a neural network. The machine listens to, say, 300 people saying ``ah'' until it learns to recognize the sound. Researchers hope that after learning these phonemes - the simplest parts of speech - the computer will know who is speaking, what language is being used, and what is being said. A simple phoneme like ``ah'' might take half an hour for the neural network to learn. More complex ones take up to two weeks.
Market predictions. According to Pittsburgh-based NeuralWare, which supplies systems that forecast financial market changes, one Texas customer used them to predict stock-index futures correctly 85 percent of the time. A major Japanese bank uses several neural networks to make market forecasts with 70 percent accuracy, the company says.
Handwriting recognition. The United States Post Office is funding research that would allow neural networks to read hand-written ZIP codes and, thus, speed up delivery.
NeuralWare's systems do everything from targeting junk mail to helping oil corporations locate drilling rigs.
How do they work? Casey Klimasauskas, president of NeuralWare, holds up a small grid of movable pins to explain.
Each pin represents a brain-like sensor or neuron, he says. As information passes to it, the neuron performs a calculation and pushes the pin out or leaves it in. As more information passes through, the network tries to make a prediction. Those neurons that were ``right'' get reinforced (the pins are pushed out further). Those that were ``wrong'' get weakened (the pins are pushed back in a little). Sometimes, the network knows what's right and wrong because a person tells it; sometimes, it corrects itse lf. Eventually, it builds a very sophisticated, three-dimensional pattern.
Neural networks use these patterns to make predictions of trends. Such methods approximate the way people think much better than so-called ``artificial intelligence'' or ``expert'' systems in which computers use a specific set of rules rather than trial-and-error to reach conclusions.
Scientists are a long way from building a neural network that begins to match the complexity of the human thought process.
``The current neural nets are very simple things,'' says Peter Greene, associate professor of computer science at the Illinois Institute of Technology in Chicago. He envisions that such electronic ``organisms'' will one day reproduce themselves.
But think like people? ``It's very hard to model that with a computer,'' Professor Greene says. ``People have such richness. There are revelations that can hold us for centuries.''