GWAP.com / CMU |
| The Squigl game involves having two players outline the same object in an online picture. Points are awarded based on how close their outlines match. |
Researchers are enlisting Internet users to try out a new set of games that will help them develop smarter search engines and sharper-eyed machines. It’s kind of like playing "Hot or Not" … for a scientific cause.
Games With a Purpose, or GWAP, is the brainchild of computer scientists at Carnegie Mellon University - including Luis von Ahn, one of the creators of the CAPTCHA filter to distinguish between real humans and machines. As part of their plan to elevate machines to the next level of human-style intelligence, von Ahn and his colleagues hope to capitalize on the all-too-human desire to get to a multiplayer game's next level.
"We're trying to do two things here," von Ahn told me today. "One is to train computers to do the things that humans do. And another thing we're trying to do is just use a different approach to experiments. For the first time, we can get hundreds of thousands or millions of people working on the same problem at the same time. ... We're seeing how we can coordinate millions of people to solve tiny bits of the problem."
Von Ahn's first GWAP offering was called the ESP Game: The same image is displayed to two players who are online simultaneously. Each player tries to guess the word that the other player would use to describe the image, and the players win points when they hit upon the same word. The goal is to complete as many matches as you can before time runs out, and get your name on the list of top scorers.
Sounds fun, right? Well, the data from the guesses can be analyzed to develop better word-to-image search engines. The idea was so good that Google licensed it and turned it into a function called Google Image Labeler.
The new GWAP Web site, which made its official debut today, adds four new games:
- Matchin, a game that asks players to judge which of two images is more appealing. The game could help developers create search engines that rank images automatically based on which ones look the best.
- Tag a Tune, in which players describe songs so that computers can search for music based on tags (for example, an upbeat song about a wedding) rather than based on the title alone.
- Verbosity, which works kind of like the old TV game "Password." You have to help your partner guess a secret word without using the word itself. The results can be used to develop better artificial-intelligence strategies.
- Squigl, a game in which players trace the outlines of objects in photographs, to help teach computers how to recognize those objects more easily. The closer your outline is to your partner's, the higher your score.
"People have been trying to work on image recognition for 40 years," von Ahn explained, "and it turned out to be a much more difficult problem than we thought it would be."
Solving the problem could lead to autonomous vehicles that would more easily identify road hazards as they drove themselves - that is, the kind of vehicle that von Ahn's colleagues at Carnegie Mellon created to win a $2 million prize from the Pentagon last year.
The technology could also make cameras smarter. "If a camera could recognize that you're not smiling, it would tell you to smile," von Ahn said. (Sony has already come out with cameras equipped with a "smile shutter.")
To take part in GWAP, you register with the site, click on a game, and then wait for an anonymous online partner to show up. It doesn't take that long, especially during U.S. daytime hours. "We just released it today, and several thousand people have already played," von Ahn said.
The data from the game play will be freely shared with other researchers working on artificial-intelligence tools as well as the psychology of human word and image recognition, von Ahn said.
The gender game
Although you're not required to share any personal data, you do have the option of filling out a profile that's much like the kind of listing you'd see on MySpace or Facebook. Von Ahn's experience with the ESP Game has already yielded some insights into the finer points of game play.
"Pairs of females and pairs of males do better than when one is female and one is male," von Ahn observed. "For example, if there's a picture of an attractive female, and there are two males playing, they're more likely to both say the word 'hot.'"
Despite that same-sex advantage, von Ahn said online players would rather go through the game with a partner of the opposite sex if they're given a choice - which just goes to show that a little sex appeal goes a long way, even when you're engaged in the scientific game.
Von Ahn's colleagues in the GWAP project are software engineers Mike Crawford and Edison Tan and graduate students Severin Hacker, Edith Law and Bryant Lee.
For more about "Hot or Not" research, check out Melissa Dahl's story about face-attractiveness studies. For more about games with a purpose, consult Kristin Kalning's "On the Level" column. And if you have any other scientific gaming suggestions, let me know in a comment below.