New games are designed to make computers smarter

Share with others:


Print Email Read Later

Computers aren't nearly as bright as movies like "2001: A Space Odyssey" would have us believe, but a quartet of free online games developed by Carnegie Mellon University scientists will help make them smarter.

Introduced today, the so-called GWAPS -- short for "games with a purpose" -- are meant to "help improve Internet image and audio searches, enhance artificial intelligence and teach computers to see," said Luis von Ahn, assistant professor of computer science and head of the team that developed the games.

"It's very strange, especially when you consider how they are portrayed in the movies," Dr. von Ahn said. "They are portrayed as someone who knows it all.

"[And] they are very good in certain things. For example, they're much better than us at arithmetic and multiplying large numbers. But for certain things like analyzing images ... if you give them an image and ask if it is a cat or a dog, they can't tell you. They can't differentiate between the cat and the dog."

Teaching such image recognition is a goal of a GWAP named Squigl, in which players trace the objects in photographs. Matchn, in which players judge which of two images is more appealing, also deals with image recognition. Tag a Tune calls for players to describe songs -- as happy or sad, for example -- so computers can search for music by a means other than title. Verbosity is designed to help computers amass facts.

Also on gwap.com is a classic developed by Dr. von Ahn called the ESP game, in which images are shown to two players who each try to guess the words the other player would use to describe the image.

A few more new games will be added soon, Dr. von Ahn said. "One of them will be about language translation. ... Computers are not yet good at translating between languages." Still another will be about classifying words, like classifying the word dog as both an animal and a pet. That, he explained, would lead to more sensible searches.

For a computer, the process of learning image recognition is not unlike the way a child learns to associate words with images, Dr. von Ahn said.

"A child sees examples and learns the word. With a computer it's the same thing. We teach them how things look. ...

"There's one big difference between kids and computers. With kids, you just have to show them five cars and they learn what they look like. With a computer, you have to give them millions of cars. ... There are different ways of doing that. One possibility is you pay people ... [or] rather than paying people to outline hundreds of thousands or millions of images, have them do it having the fun of playing a game."

Researchers also will have free access to the knowledge amassed by the GWAPs. "We'll put it on the Web and they can download it," Dr. von Ahn said.

"This is not the only avenue people are trying" for teaching computers such lessons, he said.

Other members of the development team include software engineer Mike Crawford and graduate students Severin Hacker, Edith Law, Bryant Lee and Edison Tan.


Pohla Smith can be reached at psmith@post-gazette.com or 412-263-1228.


Advertisement
Advertisement
Advertisement

You have 2 remaining free articles this month

Try unlimited digital access

If you are an existing subscriber,
link your account for free access. Start here

You’ve reached the limit of free articles this month.

To continue unlimited reading

If you are an existing subscriber,
link your account for free access. Start here