Drawing inspiration from the brain's neural circuitry, researchers at MIT and Lucent Technologies' Bell Labs have developed a computer algorithm that mimics a key aspect of intelligence: the ability to analyze complex objects into simple and meaningful components.
The research, described in the October 21 issue of Nature, provides fundamental insights into how basic properties of the brain's neurons are related to intelligent behaviors such as vision and language. It may also lead to improved videoconferencing capabilities and more powerful web-searching tools.
The researchers demonstrated the power of their algorithm by applying it to a large database of facial images. The computer learned to analyze the faces into a small set of features resembling facial parts, such as eyes, noses, ears and mouths. Then, by adding these parts together in various combinations, the computer was able to resynthesize all of the original faces.
"Using a set of parts can be an efficient way of representing highly variable objects," said Bell Labs researcher Daniel Lee. "Think about the toy, Mr. Potato Head. With a limited number of parts, a child can construct a huge variety of heads. Now think about faces. Because they are so variable in appearance, they have been difficult for computers to deal with. Our algorithm offers a new way of coping with such variability, by automatically generating a set of parts that are suitable for representing a class of objects." Lee collaborated with Sebastian Seung, assistant professor of computational neuroscience in MIT's Department of Brain and Cognitive Sciences.
Because the algorithm generates an efficient representation of facial images, it could be used for improving videoconferencing. "Instead of sending bits of information over the Internet that represent the pixels of all the faces around a table, it may be possible to send a handful of facial features that can be reassembled on the other end," Professor Seung said.
The applicability of the method is not limited to images, but extends more generally to other types of data. For instance, when the researchers applied their algorithm to a text database, it learned a set of semantic features. Each feature consisted of words with related meanings, such as "flowers," "leaves" and "plants."
Such semantic features should be useful for helping computers decide when different documents have similar underlying meanings.
"Language is tricky: two words can have the same meaning, and one word can have two meanings in different situations," said Professor Seung. "That's why web search engines do so poorly. When you ask for all documents containing a particular word, you miss the ones containing its synonyms. And you are given all sorts of documents containing the word with meanings that you didn't intend. Our algorithm could help computers learn to avoid such confusion."
A version of this article appeared in MIT Tech Talk on October 27, 1999.