sealPurdue News
_____

June 1995

Neural networks give computers capacity to learn

WEST LAFAYETTE, Ind. -- Who says you can't find good help these days?

A Purdue University researcher says computers driven by neural networks soon may serve as intelligent assistants to doctors and lawyers, professors and researchers.

The computers could perform such electronic legwork as sorting the junk out of electronic mail, fetching interesting stories from newspapers on the information superhighway, or telling a driver on an actual highway that she'll really like the restaurant that's just over the next hill. They might even perform medical diagnoses or legal research.

A neural network is a computer program, modeled roughly after the brain, that can learn to perform tasks and make decisions based on past examples or experience, says Ray Eberts, associate professor of industrial engineering.

"Any decision requiring expertise, such as medical diagnosis, legal decisions or financial investments, could be assisted by using neural nets," he says. "Just as people have procedures for learning from experience, the neural network also has learning procedures stored in the program.

"Also like people, neural nets can recognize patterns and analyze subjective information such as speech, fingerprints or stock performance better than standard programs, which are designed primarily for very precise calculations."

Neural networks "learn" by gathering and storing information from the computer user. When the computer receives new information, it uses its stored expertise to classify it or recognize a pattern.

Researchers in academia and industry have studied neural networks extensively for about 20 years, Eberts says, and some companies today use them in large-scale operations. For example, credit card companies use neural networks to analyze when spending patterns on a card change, indicating the card may have been stolen. Neural networks also are being researched for medical applications, such as diagnosis, predicting the length of hospital stays and analyzing X-rays.

While Eberts does not think the computer will ever replace the family doctor, he says it could be a valuable, time-saving assistant.

"A physician may take many years to learn how to make a diagnosis, but the neural network can store this learning on a computer file in a very small space," he says. "The flexibility to store all kinds of expertise efficiently is what makes the neural network powerful."

Some commercial applications that rely on optical character recognition also use neural networks, such as the optical readers used by the U.S. Postal Service to process addresses on mail.

Neural network software can run on personal computers, but Eberts says he's not aware of any commercial programs yet available to the consumer. He estimates neural net-based software may be developed for use in areas such as medicine, legal affairs and financial planning in the next five to 10 years, but liability questions first must be worked out.

However, Eberts says neural nets offer many other commercial possibilities.

For example, he has created a program called a neural network "agent." Eberts says other researchers also work with applications for agents. He has programmed, or "trained," an agent as a personal assistant to perform tasks such as sorting e-mail and finding journal articles on the Internet.

"I've given the agent a set of preferences that tell it what characteristics of an e-mail message I think are important and what I think is junk," Eberts explains. "Because it has learned my preferences, when a new message comes in, the agent evaluates it and sorts it according to those preferences, and it's accurate about 70 to 80 percent of the time. If I think the classification was wrong, then I can reclassify the message and the agent will relearn the classifications."

Another of Eberts' projects incorporates neural net agents into an intelligent highway vehicle system, in which drivers have a display in the car that provides information on facilities in the area, such as restaurants, motels and rest stops. The agent learns from the user the types of restaurants he or she prefers depending on things such as time of day, day of the week and number of people in the car.

"Our data indicate that a neural network can be quite accurate, 90 percent of the time or more, in choosing the preferred restaurants," Eberts says.

Neural network-based agents also can be intelligent assistants, performing tasks automatically and providing advice on how to do a task more efficiently.

"A neural net agent is different from a regular computer program because the learning process is very interactive, it 'watches' how I do things," Eberts says. "If I want to schedule a meeting, my agent should know when I am available, when I like to schedule meetings, and then correspond with the other person's agent to find a suitable time. Neural nets can do this now.

"I also have developed an intelligent neural net assistant that analyzes my computer commands. If I go through a series of commands and it is not very efficient, a window pops up on the screen that tells me how this could be done more efficiently."

Eberts' research is funded in part by a National Science Foundation Presidential Young Investigator Award.

Source: Ray Eberts, (765) 494-5429; Internet, eberts@ecn.purdue.edu
Writer: Amanda Siegfried, (765) 494-4709; Internet, amanda_siegfried@purdue.edu
Purdue News Service: (765) 494-2096; e-mail, purduenews@purdue.edu

NOTE TO JOURNALISTS: To receive the text of this news release via e-mail, send an e-mail message with the text "send punews 9505ep4" to this address: almanac@ecn.purdue.edu. Purdue News Service also maintains a searchable data base of faculty experts and posts news releases, experts lists and story tips on a web server at http://www.purdue.edu/uns and a gopher server at newsgopher.uns.purdue.edu. The web site also offers selected downloadable photographs.


* To the Purdue News and Photos Page