|
PITTSBURGH, Pennsylvania (AP) -- Computers might seem smart, but they have no common sense. Two Carnegie Mellon University researchers using a Web site called the ESP Game are among a growing number nationwide tapping into human brains for common knowledge that can be programmed into computers to improve artificial intelligence. Grad student Luis von Ahn and his mentor, computer science professor Manuel Blum, hope that search engines such as Google and Alta Vista someday will adopt word labels generated by their ESP Game to help computers see images more like the way humans do. "People talk about all this unused CPU [computer] power on the planet, but the greatest amount of untapped CPU power on the planet is the human brain," Matthew Burke, an industry analyst with New York City-based Jupiter Research said. "It's just a question of how you loop them into [computer networks] and connect them to it." Search engines use algorithms -- mathematical recipes designed to solve problems -- to sort, rank and filter pages, text and images on the Internet. But they can't "see" an image the way a human being does. Instead, the engines rely on the surrounding text to make an educated guess about what the picture shows. While that works most of the time, it's not foolproof -- and it offers no help to blind Internet users who use special equipment to read the text aloud. When such devices come upon a picture, they simply say "image" without offering any description, von Ahn said. The ESP game (www.espgame.org) works by pairing a player with an anonymous Internet partner who are both asked to type in words that describe a series of images. The players win points when they match words -- and von Ahn and Blum have another label they can affix to the image in question. It would take too long for researchers to label every one of the hundreds of millions of images that can be accessed by Google or other search engines and assign five or six word labels to each. But by getting a few thousand people to play the ESP game each day, von Ahn hopes that might be accomplished in a few months. Spokesmen for Google and Alta Vista were mum on that prospect because they jealously guard the computer secrets their programmers have developed to search the Web. Some industry analysts don't agree that efforts like von Ahn's will change the search engine landscape all that much. Sophisticated algorithms can eliminate some of those semantic problems simply by tracking which sites help the most users with specific questions -- and that's generally faster and cheaper than using a phalanx of human editors, said Danny Sullivan who edits a Web site called SearchEngineWatch.com, a branch of Connecticut-based Jupitermedia.com. "They figure, maybe we can't be 100 percent perfect every time, but we can be fairly good. Maybe we can be really good 90 to 95 percent of the time and that's good enough," Sullivan said. Google boasts that it searches more than 3.3 billion Web pages and that includes an estimated 1 billion images, Sullivan said. Still, Sullivan said some sites use human editors so that their search engines make some semantic distinctions. For example, if you type the word "Saturn" into MSN.com it suggests three broad categories: the planet, the Sega game system and the automaker. "That's in part because they had human editors go through those Web pages and say, 'We have three different meanings for the word Saturn,'" Sullivan said. Other artificial intelligence, or AI, researchers agree that computers need more human, common-sense input. "Computers don't know very much about the world -- like that clouds are fluffy, and the sky is blue, and people sleep at night. It's called common-sense knowledge; that's actually a technical term," said Push Singh, a graduate student researcher at the Massachusetts Institute of Technology. "If you buy a copy of Microsoft Windows it can do all sorts of applications, but it doesn't have any common sense in it." The Open Mind project, begun three years ago (openmind.media.mit.edu) by Singh and others at MIT, solicits common-sense information from visitors by asking them to supply underlying "common-sense" facts that fit a given scenario. Users are asked to list five things that someone would need to know beforehand to understand a statement like "Bob bought some milk." Three examples are, "Bob probably used money to buy the milk;" "Milk is a kind of food;" and "Bob probably bought the milk at a store." It's not rocket science, but such common sense is beyond even the most sophisticated computers. For now. "If we can crack the common sense problem, then we've solved AI," Singh said. "Figuring out the common sense problem is almost the same as being able to build a person." |
________________________________
Changes to your subscription (unsubs, nomail, digest) can be made by going to http://sandboxmail.net/mailman/listinfo/sndbox_sandboxmail.net
