WIRED MAGAZINE: ISSUE 15.07
For Certain Tasks, the Cortex Still Beats the CPU
Which is prettier? A picture of a black cat sleeping on a pillow or one of a curly-haired brunette woman in a miniskirt? I've got only a few seconds to decide. I vote for the cat. I'm sitting in a laboratory at Carnegie Mellon University, playing Matchin', a computer game developed by Luis von Ahn. In the game, two players — von Ahn and I, seated at different terminals — watch as pairs of pictures swiped off the Internet flash up on our screens. Our goal is to pick the one we think both of us will find more attractive, not necessarily the one we personally prefer. This requires a sort of mindmeld, and it doesn't always work: Von Ahn picks the girl in the miniskirt instead of the cat. We've got one minute to process as many pictures as we can, so we race on frantically, evaluating photos in an instant. Soon we hit a groove: We both say that a picture of a peacock is prettier than one of a picnic, a baby is lovelier than a tombstone, a wedding couple beats a field of wheat. Then the game is suddenly over, and we get our score: We agreed 70 percent of the time. Pretty good, but not enough to hit the high-score tables.
"Man," laughs von Ahn. "You picked some weird stuff!"
Games that Help Solve Computing Problems. Your Turn.
Luis von Ahn's new games pair random players to solve a computing problem. Because the two players get points when their answers match, the accuracy — and fun quotient — increases. To try them, go to www.gwap.com.
1) Matchin' Players are shown the same pair of images, then each tries to pick the one they'll both agree is more attractive. Creates a database of images searchable by aesthetic value, a task no algorithm can perform.
2) Babble Two English-speaking players are shown a sentence in a foreign language that neither of them speak. A list of possible English meanings appears below each word. Players try to agree upon a set of English words that forms the most coherent sentence. Translates foreign text into English without requiring anyone fluent in both languages.
3) InTune Players listen to the same audioclip and then try to come up with the same phrase to characterize it. Tags sounds with searchable descriptive text.
4) Squigl Two players are shown the same picture and a word describing an element within the image (e.g., a picture of a dog and the word "leash"). They each draw a border around the element. Produces a set of pictures with their internal components tagged — terrific for very specific image searches.
5) Verbosity One player is given a word, and the other tries to guess that word by completing phrases such as "It is near a ____" or "It is a type of ____." The first player answers "true" or "false" but can't use the word itself. Creates a database of commonsense knowledge describing the objects.
It's an oddly enjoyable game. But Matchin' is also a covert experiment in artificial intelligence. Every time players agree on a picture, it's tagged as prettier. Von Ahn, a 28-year-old professor of computer science at Carnegie Mellon, will put the game online this summer, and as thousands of people play it, his database of 100,000 photos will be imbued with something quintessentially human: an aesthetic sensibility, encoded as a ranking of attractiveness.
The game basically tricks humans into teaching computers what constitutes prettiness. If enough people play Matchin' — and von Ahn's previous games have garnered millions of play-hours — it could eventually rate the appeal of every image on the Internet. Google could incorporate the ratings into its search engine, so you could search specifically for "beautiful" pictures of houses, people, or landscapes.
"People are good at figuring out what's attractive, and computers are good at quickly searching and finding," von Ahn says. "You put them together, and bang!"
This is "human computation," the art of using massive groups of networked human minds to solve problems that computers cannot. Ask a machine to point to a picture of a bird or pick out a particular voice in a crowd, and it usually fails. But even the most dim-witted human can do this easily. Von Ahn has realized that our normal view of the human-computer relationship can be inverted. Most of us assume computers make people smarter. He sees people as a way to make computers smarter.
Odds are you've already benefited from von Ahn's work. Like when you type in one of those stretched and skewed words before getting access to a Yahoo email account or the Ticketmaster store. That's a Captcha, which von Ahn developed in 2000 to thwart spambots. Or there's von Ahn's picture-labeling games, which have lured thousands of bored Web surfers into tagging 300,000 photos online — doing it so effectively that Google bought his idea last year to improve its Image Search engine.
Last winter, von Ahn was awarded a $500,000 MacArthur genius grant, and in April he received another $200,000 as one of Microsoft's New Faculty fellows. This summer he's putting the money to good use, launching five new games that will identify sounds, give computers commonsense logic, and even help scanners perfect their optical character recognition.
"Captchas would be enough for most people to build their entire career on," marvels Josh Benaloh, a cryptographer at Microsoft who hired von Ahn as a summer intern just three years ago. "Luis keeps coming up with new stuff."