November 4, 1999

WHAT'S NEXT

A Digital Brain Makes Connections

By ANNE EISENBERG

he ability to recognize patterns -- the way noses tend to appear in the middle of faces, how words like "throw" are often associated with "curve ball" and "touchdown pass" -- has always been linked with human intelligence. Not necessarily high intelligence, but something that machines have a hard time with.

	Mary Ann Smith

In the process of investigating how the brain might accomplish such impressive feats, two physicists have devised a novel program that can recognize the existence of patterns, although it cannot yet figure out what they mean.

Their creation, a learning algorithm, may one day inspire not only better searches on the Web, but also wide-ranging improvements in computer applications like voice recognition and medical imaging and other applications that call for a computer to parse a mass of complex data.

The algorithm, a math-based strategy, was developed by Sebastian Seung, an assistant professor in computational neuroscience at the Massachusetts Institute of Technology, and Daniel Lee, a researcher in the biological computation department at Bell Labs, the research and development arm of Lucent Technologies in Murray Hill, N.J. Their account appears in the Oct. 21 issue of the journal Nature.

"Our method seeks to imitate the way in which we think the brain uses patterns to make sense of the world," Dr. Seung said.

The two physicists, part of the community of scientists working on the border of artificial intelligence and neuroscience, tested their learning algorithm with images and with words. For images, they gave it pictures of faces. For words, the algorithm was provided with encyclopedia articles. The job was the same: to find recurring pieces of text or images and put them into groups.

The algorithm works by the statistical analysis of words or pixels that occur together and that may therefore be related. For faces, the algorithm looked for pixels that repeatedly popped up together.

"It was not looking specifically for noses but for repeated patterns," Dr. Seung said. "We didn't explain about noses." On its own the program came up with groups that quite visibly resembled noses, mouths or other features constituting identifiable facial parts.

For text, the algorithm grouped words like flower, leaves, plant and perennial together. Once the groups were created, the researchers labeled them as belonging to a topic, like botany.

The method for identifying images of facial parts or topics in text relied on using the algorithm to analyze huge amounts of data: volumes of an encyclopedia and thousands of faces.

"It's surprising what you can discover about meaning from statistics, that is, essentially from what gets used together," Dr. Seung said.

A computer is equipped to find recurring patterns in text or images.

Dr. Lee emphasized that the researchers' learning algorithm worked without supervision. They did not tell the algorithm which features constituted a face, for instance, or where to look for items that might belong in the category "botany." "Much of human learning is done without a teacher around," Dr. Lee said. "We wanted an algorithm that could in an unsupervised fashion recognize correlations."

An artificial-intelligence expert, Geoffrey Hinton, director of the Gatsby Computational Neuroscience Unit of University College, London, said Dr. Seung and Dr. Lee's approach was an important contribution.

"It's a novel way of taking large data sets and automatically finding features," Dr. Hinton said.

The new approach may also lead to valuable information on how the brain extracts features from images, Dr. Hinton said. People can become very good at doing that, he said, using as an example someone who was an expert reader of X-rays.

"People learn over several years to see details in X-rays that they could not initially perceive," Dr. Hinton said. But it is not clear how the brain works to increase this skill.

"When we understand how the brain extracts different features," Dr. Hinton added, "we can make computer vision systems that are better at interpreting medical images."

At AT&T Laboratories in Florham Park, N.J., Dr. Fernando C. N. Pereira, head of machine learning and information retrieval research, confronts problems related not to how the brain processes images but to how it handles language. "This is an important advance," he said of the new approach. "It brings together two seemingly separate areas -- images and language -- and shows that they can be subsumed under one single model."

Dr. Pereira, an expert in language processing, said one of the most challenging problems in the field was determining how people learn certain cues that help them understand language.

"Children learning language may draw from words and sentences in context to get meaning, plus they also have other cues from seeing and hearing," he said.

Computers, in contrast, do not yet have this rich sensory capability to draw on when confronted with the language used in daily life.

Like Dr. Lee and Dr. Seung, Dr. Pereira analyzes acres of text, using statistical methods to find indicators of the meaning that words have in context.

"As we build better models of how the underlying structures of speech and text are learned," he said, "we can apply them to speech recognition and information retrieval."

Dr. Pereira avoids the word "understand" when he talks about the models he is developing that help computers handle human speech and text.

"My work is in having computers process and react appropriately to natural-language speech and text," he said. "The term 'understanding' turns out to be too controversial. I'm settling for 'react more appropriately.' "

What's Next is published on Thursdays in the Circuits section. Click here for a list of links to other columns in the series.