August 15, 2006
[ return to the news list ]
Technology Review (05/25/06) Savage, Neil
Researchers at MIT's Center for Biological and Computational Learning (CBCL) are attempting to improve computers' ability to discern the properties of images. Visual searching is nowhere near as accurate and thorough as text-based searching, though if the researchers are successful, their work could eliminate the need for humans to monitor surveillance cameras at a military base or an office, for instance, though they admit that their work will not be easy. "The fact that it seems so easy for a human to do is part of our greatest illusion," said Stanley Bileschi, who was recently awarded a PhD in electrical engineering and computer science at the CBCL. Interpreting visual data is a complex task for computers, notes Bileschi, adding that humans exert roughly 40 percent of their brains on that single task. The multiple variables that factor into identifying visual data include color, lighting, spatial orientation, distance, and texture. Traditional image-recognition systems rely on statistical learning systems to train computers to identify objects by examining each pixel in the image. Neurophysiologists at the CBCL are attempting to improve on this method by exploring how the human brain processes images, pointing to the way that a photoreceptor in the eye is stimulated by each pixel. The stimulus causes neurons to fire in a specific pattern, which the researchers track with mathematical models and train the computer to associate each neuron simulation pattern with its corresponding object. Just as a baby learns to distinguish trees from faces, the computer remembers the pattern when it encounters the same object again. Tests have shown that the computers can identify people and cars in a street scene with an accuracy rate between 95 percent and 98 percent.
For the complete article, see http://www.technologyreview.com/read_article.aspx?id=16926&ch=infotech
Posted by rab5 at 09:42 PM
[ return to the news list ]
