MURRAY HILL, N.J. -- Drawing inspiration from the brain's neural circuitry, researchers at Lucent Technologies' Bell Labs and the Massachusetts Institute of Technology have developed the world's first computer algorithm that mimics a key aspect of intelligence: recognizing meaningful patterns in large collections of images, text or other data.
The research, described in the current issue of the journal Nature, provides fundamental insights into how basic properties of the brain's neurons are related to intelligent behaviors, such as vision and language. It also may lead to improved Web searches, videoconferencing, and data storage and transmission.
"This program learns about the world in much the same way that people would -- by looking for elements that repeatedly appear in a group of objects," says Bell Labs researcher Daniel Lee. For example, when presented with 2,400 photographs of human faces, the program figured out that all of the faces contained common elements -- such as eyes, noses, ears, and mouths -- and then looked for similarities among those features. The program then was able to use these 49 features to recreate all 2,400 faces.
Lee collaborated with Sebastian Seung, MIT assistant professor of computational neurosciences in the department of brain and cognitive sciences, on the program, which is an algorithm for non-negative matrix factorization.
The algorithm can similarly analyze other collections of images or documents to discover patterns. For example, when the researchers applied the program to an encyclopedia database, it was able to group the thousands of articles into 200 general topics based on their content. One such topic was related to botany, for which the algorithm automatically learned to group words together such as "flowers," "leaves," and "plants."
Because the algorithm finds an efficient way of representing data, it could be used in the future to compress video, images or other data for storage or transmission. Applied to Internet searches, the program could find relevant Web pages that do not necessarily contain the specified key word and also could screen out pages that contain the search term, but in the wrong context. For instance, it can learn to distinguish documents dealing with the metal "lead" from articles about leadership.
"Our algorithm could help computers learn to avoid such confusion," Seung said.
Lucent Technologies, headquartered in Murray Hill, N.J., designs, builds and delivers a wide range of public and private networks, communications systems and software, data networking systems, business telephone systems and microelectronics components. Bell Labs is the research and development arm for the company. For more information on Lucent Technologies, visit the company's web site at http://www.lucent.com or the Bell Labs web site at http://www.bell-labs.com.
