COMPUTERS THAT LEARN BY DOING Programs and chips that mimic the way the brain works are catching on in business. They spot credit card crooks, pick stocks, sort apples, and even drive cars.
By Gene Bylinsky ASSOCIATE REPORTERS Alicia Hills Moore, Jane Furth

(FORTUNE Magazine) – IN A SUITE of small offices strewn with electronic equipment in an old building in downtown Palo Alto, California, the ten boyish-looking Ph.D.s who constitute Lexicus Corp. are making computer history. Dressed in shirtsleeves and shorts on a Saturday morning, Ronjon Nag, 31, CEO of the 18-month-old company, scribbles with a stylus on the tablet-like screen of a notepad computer: This is a demonstration of Longhand. The machine, about the size of an 8 1/2-by-11-inch writing tablet and 2 1/2 inches thick, prints his words on its screen instantly and accurately. To be sure, the British-born, Cambridge- educated Nag (rhymes with log) leaves more space than usual between words and writes them somewhat more carefully than he normally would. Still, in a world where computer industry wags contend that the acronym PDA (personal digital assistant) -- another name for some notepad computers -- really means ''product deficiencies abound,'' the wonder is that the thing performs one of the most complex tasks in computing reasonably well.

The inability to recognize natural handwriting has slowed the acceptance of so-called pen-based computers to a crawl. The few such machines on the market so far can read only hand-printed letters, and even those with varying degrees of accuracy. Lexicus is about to change that. You can already buy its Longhand program on a floppy disk for $495 to use with existing notepads such as the AT&T/EO, Grid Convertible, Toshiba Dynapad, or NEC VersaPad. Notepads to be introduced this year by Compaq and others could begin to justify some of the hype that greeted the appearance of handwriting recognition technology a few years ago. Notepads with Longhand will compete against Apple Computer's much publicized Newton, which just entered the market. Newton can also recognize natural handwriting -- but only about half as quickly as Longhand can. In a fast machine like the AT&T/EO, the Lexicus program identifies three short words per second. The big reason Lexicus gained a jump on the market: The developers of Longhand took advantage of their familiarity with neural networks, a new approach to computing that roughly mimics the brain's unmatched ability to recognize and understand patterns -- faces, voices, and, in this case, written characters. ''Neural'' comes from neurons, or brain cells. Computer scientists build neural networks by imitating in software or silicon the structure of brain cells and the three-dimensional lattice of connections among them. By contrast, Apple's Newton relies on the older, less sophisticated techniques of classical mathematical algorithms, or formulas, for handwriting recognition. But Apple too plans to use neural networks later this year in a Macintosh notepad that will recognize hand printing -- but not script. The hot news about neural networks is that after years of development and sporadic trials they are finally hitting the big time -- and giving their users an edge in products and services. Neural computation is fast gaining ground among both software and chip developers. In a recent survey, the trade paper Electronic Engineering Times found that 85% of the engineers it questioned in the U.S., Europe, and Japan ranked neural computation as the hottest emerging computer technology. ; The neural network business is already booming. Says Robert L. North, CEO of HNC Inc. of San Diego, the largest supplier of neural network hardware and software: ''Neural networks have finally moved from a laboratory curiosity into everyday usage in a wide variety of industries.'' Those applications now cover a surprising range: sorting good apples from bad, for example, or spotting bad apples of another variety -- fraudulent users of credit cards -- with unprecedented precision. They help manage big investment funds, outperforming Standard & Poor's 500-stock index in the process; they detect signs that industrial machinery is about to fail, alerting workers to the need for preventive maintenance. In every instance, users are taking advantage of the networks' remarkable ability to discern patterns and trends too subtle or too complex for humans, much less conventional computer programs, to spot. While people can't easily assimilate more than two or three variables at once, neural networks can perceive correlations among hundreds of variables. Neural nets shine precisely in areas where conventional computers are dim: performing many operations simultaneously, recognizing patterns, making associations, generalizing about problems they haven't been exposed to before, and learning by experience. Neural networks sometimes perceive trends that people have overlooked -- new types of credit card scams, for instance. Late last year, using software from Nestor Inc., a Providence neural network producer, Intel introduced a neural net chip that contains more transistors than its latest supermicroprocessor, the Pentium. Motorola, National Semiconductor, and other chipmakers are all working on neural designs. Over the next five years, the chips will begin to bring the power of these silicon brains not only to your PC but also to such mundane yet technically tricky tasks as automatically balancing shifting laundry loads in washing machines and telling microwave ovens how long to cook a roast or chicken. ''You'll see neural network chips in all computers,'' says Joshua Alspector, director of neural network studies at Bellcore, the research arm of the Baby Bells. ''They'll serve as information-filtering tools on your PC. They'll know your preferences and adjust to your needs. They'll sort your E-mail in order of priority and select articles from the AP news wire to store for you. They'll simplify the use of computers.'' Ordinary computers mechanically obey instructions written with uncompromising precision according to set rules. The most startling thing about neural networks is that, like children, they learn by example. Think of a neural network as a system of pipes with valves that control the flow of water from one reservoir -- or neuron -- to another. The network learns by fine-tuning these valves in response to each situation it encounters. Whether you're training a neural network to speak English -- as Terrence Sejnowski, a neuroscientist now at the Salk Institute, and his graduate students did in a classic demonstration a few years ago -- or to distinguish bad apples from good, the principle is the same. In the brain, learning takes place through changes in the links between neurons. If you're teaching the neural network to speak, for instance, you train it by giving it sample words and sentences, as well as desired pronunciations. The connections between the electronic neurons gradually change, allowing more or less current to pass. Electronic neurons are extremely primitive representations of real ones. But you wouldn't know it listening to the tape of Sejnowski's network, which learned to pronounce a 20,000-word vocabulary overnight. The network sounds uncannily like a child, at first groping for the right phonemes, then gradually mastering words and sentences. TO TEACH a computer to tell good and bad apples apart, you show examples of each to a color TV camera connected to a neural network and tell the network which are which. Once you have shown it enough good and bad apples, the neural network can direct a mechanical device to toss aside the bad ones. No magic is at work. The networks cannot function well without high-quality data; the more examples they are shown the better. Training a neural network remains something of an art, however. (To master the techniques, business users often find it helpful to take the weeklong courses offered by companies such as NeuralWare of Pittsburgh, a privately held supplier of neural networks.) Singular, unexpected events can stump a network. A neural system trained to trade stocks, for instance, wouldn't have known how to respond, say, to the outbreak of war in the Gulf unless it had been given examples of how the market had reacted to earlier wars. Human judgment is still necessary to keep the networks running properly. Even so, used judiciously, neural networks have an uncanny ability to spot emerging trends and often draw correct conclusions better than people do. Here's how smart users in a number of fields are gaining on competitors by applying neural networks:

-- Detecting credit card fraud. Banks and other card issuers have been losing the battle against stolen and counterfeit cards. According to a recent Nilson Report, an industry newsletter, credit card issuers' losses worldwide are expected to reach $2.4 billion this year, triple the 1989 amount. Until recently the standard approach has been to use expert systems -- computer programs that incorporate knowledge painstakingly assembled by asking specialists in specific areas of expertise how they do their work. In fraud detection, such programs usually look for sudden, obvious changes in a cardholder's spending patterns -- expensive clothing or jewelry purchases, say, or large cash withdrawals. Expert systems flag far too many card users who may have altered their spending only slightly. Result: Card issuers wind up calling lots of innocent cardholders, irritating honest customers and wasting limited resources. Neural networks sharply narrow the number of suspects. At Mellon Bank's Visa and MasterCard operation in Wilmington, Delaware, which keeps track of 1.2 million accounts, the expert system looked at only a few factors, such as the size of the transaction. The computer fingered as many as 1,000 potential defrauders a day. Mellon's fraud investigators couldn't keep up. Says Philip J. Samson, a consultant who was until recently a Mellon vice president: ''The system almost became useless.'' Since Mellon Bank put in a neural network early in 1993, it deals with only one-tenth as many suspicious transactions. ''The neural network does half of my job,'' says fraud investigator Cynthia Ciamaricone. She and her colleagues now apply their judgment to fewer cases faster and more effectively. With the expert system, investigators usually didn't get around to checking on a transaction for a couple of days. Now they do it in less than two hours. In one case, the neural net allowed Mellon Bank to notify a woman that her credit card had been stolen even before she missed it; the theft had happened only hours earlier. Says Barbara Donovan-Lamb, the Mellon fraud detection manager: ''This is making finding that needle in the haystack a lot easier.'' The neural network also spots trends even before bank officials do. Samson recalls that one morning last January, he had just learned from another bank about the so-called California scam. Card thieves and counterfeiters had begun using a new trick to test whether a card had been reported stolen: They started doing $1 transactions, usually at gas stations. Later that day, when Samson went to check a workstation screen, he was stunned to see the just- installed neural network printing out a whole series of $1 transactions. ''The neural network detected this new pattern on its own, without us having told it anything about the scam,'' marvels Samson. Mellon Bank paid Nestor Inc. something under $1 million for the fraud- detection software. Samson figures it will have paid for itself after six months. Other banks are enlisting neural nets to combat fraud. Among them: First USA Bank's credit card division, also in Wilmington; Colonial National Bank of Horsham, Pennsylvania; and Eurocard Nederland of Rotterdam.

-- Optical character recognition. In addition to reading handwriting, which they do electronically by responding to the pressure of a stylus, neural networks have lately been joined to standard large optical character recognition systems to improve their eyesight. Those older systems, used for automated processing of forms and documents, were usually limited to reading typed or clearly printed numbers and block letters. Neural networks improve accuracy and recognize sloppier writing. Signet Bank, a major credit card issuer in Glen Allen, Virginia, uses a neural network to process automatically student loan applications and about 35% of the one million canceled checks it handles each day. Neural nets can handle all sorts of forms by reading the information directly and sending it into corporate data banks. David Fox, CEO of Nestor, sees this as the beginning of the end of keyboarding data in low-cost-labor locales like Ireland and India. Scott Blau, CEO of Datacap Inc. of Tarrytown, New York, which tailors neural nets for processing corporate forms, says revenues have been doubling each year since 1990; he already has 150 installations in a variety of industries. Says Blau: ''It's dawning on the world at large that it can automate data entry with this technology.''

-- Industrial applications. Using neural network software from Nestor, researchers at IBM France have built Neuroscope, a diagnostic tool that runs on a PC and provides early warning of failure in industrial machinery such as motors, cleaning tools, and pneumatic robots. Neuroscope makes possible ''predictive maintenance,'' spotting incipient problems days earlier than expert systems can. ''It's very difficult to write an expert system,'' says Jean Yves Leclere, who helped develop Neuroscope. ''The signs of normal and abnormal operation of machines are difficult to describe. We teach neural networks, for example, by having them listen through a microphone to the normal and abnormal sounds of a motor.'' IBM has installed Neuroscope on some of its production lines in France and markets the software in Europe. U.S. introduction is planned later this year.

-- Trading stocks. In Boston, Bradford Lewis, who runs Fidelity's Disciplined Equity Fund with a neural network, has consistently beaten the S&P 500 -- last year by 5.4 percentage points. His network usually picks little-known stocks by looking for specific patterns. For example, the neural network tries to calculate how the price of a stock may be affected by changing patterns in debt level, cash flow, earnings estimates, and other variables. For Deere & Co. in Moline, Illinois, James Hall, an engineering Ph.D. and manager of investment analysis, operates a $100 million company pension fund with a neural network, again outperforming the common indicators. Once, says Hall, his boss overruled the neural network when it suggested the fund should invest 40% of its assets in bank stocks. The boss turned out to be wrong. At LBS Capital Management, a fast-growing firm in Safety Harbor, Florida, Dean S. Barr runs the company's two funds, totaling $300 million, with a hybrid expert systems-neural network program. He too has beaten the S&P consistently, in 1992 by two and five percentage points respectively. ''The network is the ultimate decision-maker,'' says Barr. ''It does a better job than most analysts can and does it more consistently because it is not emotionally involved in the decisions.'' Again and again, Barr adds, the neural net has made timely suggestions to sell stocks weeks before their prices started to plummet. You can try neural net stock trading on your PC with such programs as the $1,295 Stock Prophet from Future Wave Software of Redondo Beach, California. (To operate in the neural network mode, the program also requires BrainMaker from California Scientific Software, which comes in two versions for $195 and $795.) Ronald Ogren, an aerospace engineer, developed the Stock Prophet because he was unhappy with his broker's performance. As with other neural network applications, stock trading requires good data -- lots of good data. In the Stock Prophet, Ogren normally uses 15 variables to predict the price trend of a stock. These include the stock's price and trading volume as well as futures contracts, the price of gold, and short- and long-term interest rates. -- Improving real estate appraisals. Foster Ousley Conley, a fast-growing company in Walnut Creek, California, has developed a neural network-based system for residential real estate appraisals. The system is now in use in a number of Western states and is coming to New York and New Jersey soon. Michael E. Foster, the company's CEO, says his company is essentially introducing quality control in a service industry. Neural nets look at the usual variables -- location, interior layout, bathroom wall materials, and so forth. But the nets do real estate valuations more precisely than human assessors because the nets compare the data with those from hundreds of other houses and analyze the variables in many different ways. Says Foster: ''We're seeing the ultimate in objectivity -- a statistically correct appraisal.''

-- Making transport more efficient. Speeding along an isolated highway at 55 mph with no one in the driver's seat, Alvinn (for autonomous land vehicle in a neural network) is an eerie sight. The steering wheel turns all by itself. Scientists at Carnegie Mellon University in Pittsburgh reconfigured an Army ambulance and trained it to drive on dirt roads as well as cross-country, avoiding obstacles as it goes along. Potential applications: unmanned military vehicles exposed to enemy fire or civilian trucks navigating around a hazardous waste site. Smart cruise control for your car could sense when the driver is nodding off, awaken him gently, and steer the car back onto the road. ''We've passed the stage of the usual university project -- just generating ideas,'' says Charles Thorpe, who heads the program. ''We're generating real prototypes that work.'' Airlines in the U.S., Britain, and Australia are beginning to use neural networks to forecast load factors and revenues by predicting how many passengers will be on a given plane and what fares they will be paying. The neural networks anticipate demand based on time of day, day of the week, and season. The networks have proved up to 20% more accurate than predictions derived from more static formulas. USAir, the biggest domestic airline so far to use neural nets, got its system from BehavHeuristics, a small company in College Park, Maryland.

Where are neural nets going? Today's are rather primitive and infinitesimally small compared to the human brain. About 100 billion neurons power the brain; even the densest electronic neural networks contain at best a few hundred rough approximations of real-life brain cells. The brain of a mere slug has 20,000 neurons. Real neurons also have many more interconnections to each adjacent neuron -- anywhere from 1,000 to 50,000. Neurons use not only electricity but also chemicals to communicate. One thing is certain. Neural networks will soon start working with blinding speed, which may help compensate for their simplicity. Today most neural networks take the form of mathematical simulations embedded in software that runs on ordinary microprocessors. The advent of neural network chips -- neurons physically copied in silicon -- could speed their operation as much as ten thousandfold. In his lab at Bellcore in Morristown, New Jersey, Joshua Alspector runs two Sun workstations side by side. One is driven by a software program and the other by four neural network chips. A geometric figure rotates on both screens, but the one on the workstation with neural network chips spins a hundred times faster. Hinting at what else neural net chips can do, scientists at the University of California at Berkeley have designed a chip that they say can mimic such animal attributes as a cat's night vision or the sonar sensitivity of a bat. The U.S. Navy already uses neural networks to identify submarines underwater by listening to the sounds their engines make. The Berkeley chip could allow video cameras to edit an image automatically to improve contrast or color, and to convert pictures instantly to halftone images for printing. Two arrays of the chips in tandem could serve as a robot's eyes, giving it depth perception and thus the ability to gauge distances. A bionic eye could eventually replace missing human eyes. NEURAL NETS are still in their infancy, but some are already Wunderkinder. One example: a self-training robot that swings from bar to bar in the lab much as a monkey or a gibbon swings from branch to branch in the jungle. It can take as many as 200 tries for the robot to learn a new routine. As it learns, the robot often falls ignominiously, but once it masters the complex underhand and overhand moves through trial and error, it executes them with panache. The robot's developer, Toshio Fukuda of Nagoya University in Japan, says that one day such self-teaching robots might serve as workers on skyscraper construction sites, inside hazardous plants, and anywhere else that's dangerous. Says Edward Rosenfeld, publisher of an industry newsletter, Intelligence: The Future of Computing: ''Neural networks provide an adaptive, human-style intelligence that will enable future computers to fly planes, run factories, hear, speak, see, understand, and discover.'' Well, perhaps. But maybe someday Mr. Data, the android wizard with electronic neurons flashing inside his head, will be more than just a creature of the Star Trek creators' imagination.

BOX: PICKING THE PONIES

Neural networks can pay off in all sorts of unexpected ways. Using a PC and a $195 neural net program from California Scientific Software of Nevada City, California, Don Emmons, a professional bettor, picked the winning horse in 17 out of 22 thoroughbred races at the Detroit course in 1990. He gathered data on the various horses' past performance and fed the numbers into the program.

SAVING ON JURORS The Montgomery County Court House in Norristown, Pennsylvania, uses a neural network to save $40,000 a year in fees for unneeded jurors. With two previous years' data on juror utilization and the next day's trial schedule, the system provides an immediate readout of the number of jurors required. The jurors, alerted 30 days ahead of time, are called -- and paid -- only if they are needed.