When you do an image search on the web, you might find what you're looking for. Those searches use captions and other text around pictures to give you results. But what if a computer could recognize a horse because it was shaped like a horse? That's what a new program called LEVAN can do.
LEVAN teaches itself and then shows you everything it knows about horses. The program was developed by UW computer scientist Ali Farhadi, and his colleagues Santosh Divvala of the Allen Institute of Artificial Intelligence and Carlos Guestrin, a University of Washington professor of computer science and engineering.
Ali Farhadi says LEVAN uses images to learn concepts, then shows you everything it learns.
"Basically when a Query hits LEVAN," says Farhadi, "it crawls all over Google Books, to figure out in language, what are the terms that could modify a concept. Then we hit image search engines like Google images or Bing Images to download images for each of those variations." Farhadi says LEVAN has an algorithm that it uses the downloaded images and teach itself the concept. You can try it for yourself, or even submit new concepts it hasn't learned yet. Editor's note: An earlier version of this story incorrectly stated Professor Carlos Guestrin's affiliation with the Allen Institute. Guestrin is a professor of computer science and engineering at the University of Washington.