Neural networks at the crossroads.

357 T HE field of neural networks was known to only a small group of scientists just a decade ago, but after the development of the back-propagation algorithm in 1984, the field cxperienced incredible growth. By early 1989, the Joint Conference on Neural Networks held in San Diego was more of a religious experience than a scientific conference. With the press looking on, the 1,600 true believers in attendance were inspired by the glow of the artificial neuron. It was as though a new and mysterious force had been discovered, and headlines in Newsweek, the New York Times, and the Wall Street Journal conveyed the notion that computers had just crawled out of the primordial ooze. Perhaps they had. The field has grown considerably since. The earlier giddiness and hype have given way to a cautious optimism fueled by a number of successful applications in a variety of applied fields. Neural networks have been successfully used in consumer products such as 35-mm cameras to dampen motion and in washing machines to clean clothes better. They have been used to pilot fighter planes (in simulations), identify enemy tanks, control manufacturing, evaluate credit applications, predict stock market trends, forecast the weather, and inspect poultry. Not surprising, neural networks have also been applied to medical diagnoses. Neural networks have been applied to two general areas of radiologic diagnosis, which is consistent with the