ISU researchers make artificial neural network discovery

Idaho State University computer science researchers recently discovered a new algorithm to help train artificial neural networks used by industries.

By Courtney Cobb – Journal Writer

What’s that mean?

Artificial neural networks essentially operate like the nervous systems of a human or animal and have artificial neurons inside.

The networks are designed much like computers, but whereas a person must tell a computer how to solve a problem, an artificial neural network can solve the problem on its own.

"(The neural network) learns based on what’s shown to do and can approximate any function," says Vitit Kantabutra, ISU associate professor of computer science.

Kantabutra says the artificial neural networks are used to recognize images, handwriting, and voice.

The networks are used in industry especially when dealing with character recognition.

Kantabutra says some scanners in today’s market have artificial neural networks and when a program is activated it can read the document and recognize it.

Another application for the networks is weather forecasting and predictions.

For the artificial networks to function, they have to go through a series of training exercises where they might repeatedly see various concepts until they can recognize the real answer.

For example, a network might be shown 2,000 variations of the letter A and after its training will be able to determine if a letter is or is not A.

One of the problems with training the neural networks is that in different situations a network might not stop training, but continue the cycle, Kantabutra says.

This is one of the reasons why neural networks aren’t widely used.

"The most important of these problems is that these networks have to be trained to learn to solve the problems they are meant to solve, much like humans and animals have to be trained to solve problems," Kantabutra says.

He says training algorithms are often slow and can get stuck indefinitely including fast training algorithms.

Over the past year Kantabutra and ISU computer science students Elena Zheleva (now a graduate student at the University of Vermont), Batsukh Tsendajv (a computer science undergraduate), Angela Hillier (computer science major and graduate biology student) and Steven Miller (computer science undergraduate) tried to develop an algorithm that wouldn’t get stuck during the training process.

Kantabutra says the researchers focused on three different problems. The first two were highly technical and benchmark problems or popular problems used to measure the training of an algorithm.

The third problem revolved around character recognition with the numbers zero through nine.

The new algorithm researchers discovered, when tested on some well-known difficult problems, trains neural networks quickly and reliably without getting stuck.

Kantabutra says he hopes the new discovery extends to another class of procedures for other neural networks.

"There are a lot of exciting things going on with the world of computers today including the Internet and programming," he says.

The research project is funded by a grant from the Experimental Program to Stimulate Competitive Research in Idaho and the National Science Foundation.

http://www.journalnet.com/articles/2003/12/01/news/local/news08.txt