TOWARD UNDERSTANDING HOW WE LEARN
To understand intelligence, one must understand the process of learning. Only then are you able to build machines that can perform functions typical of the brain, advises Tomaso Poggio, the Uncas and Helen Whitaker Professor in the Department of Brain and Cognitive Sciences.
To that end, Professor Poggio, an Artificial Intelligence Laboratory affiliate and co-director of the Center for Biological and Computational Learning (CBCL), and colleagues are studying the problem of learning at the three levels of theory: engineering, implementation and neuroscience. The application level is important to show that the algorithms derived from the mathematics will work in real-life situations. The neuroscience level gives useful suggestions and insights for the theory and engineering of learning since the brain routinely solves very difficult learning problems, such as learning to recognize objects and faces.
For Professor Poggio, this has a broad range of applications, including the development of a new class of search engines for images, video and multimedia databases. "The basic metaphor we're using in our approach is one of learning from example, using a machine that you can train just by presenting a few examples of paired inputs and outputs -- images and labels, for example," he said.
CBCL sponsors are the Office of Naval Research, the Defense Advanced Research Projects Agency, the National Science Foundation, Honda, Kodak, Compaq, the Central Research Institute of Electric Power Industry, NEC, AT&T, NTT, Merrill-Lynch, Siemens and Daimler.
This column features summaries of MIT research drawn from several sources. If you have an item to suggest, send it to Elizabeth Thomson, News Office assistant director for science and engineering news, Rm 5-111, or email@example.com
A version of this article appeared in MIT Tech Talk on August 9, 2000.