Google licenses professor's technology

On September 1, Google introduced the Google Image Labeler, an online game that invites players to label untagged images in the Google Image Search database. Carnegie Mellon students may find the game strangely familiar; that’s because it was developed by Luis von Ahn, an assistant professor in the computer science department and the creator of the ESP Game, which he licensed to Google earlier this year.

Von Ahn specializes in “human computation,” which he describes as “the channeling of humans to train and develop computers.” For his doctoral and post-doctoral work at Carnegie Mellon, von Ahn and his colleagues developed the ESP Game to help computers “learn” to identify images and words.

“You learn to see because your parents tell you, ‘that’s a dog,’ ” von Ahn said. “But nobody bothers to teach computers.” That’s where humans play an important role.

The model of the ESP Game and the Google Image Labeler is simple: A player is paired with a random partner online and presented with a series of unnamed images. Each partner types in as many accurate labels as possible without knowing what his partner is typing.

When the two partners both type in the same word, a match is made and the image is labeled with that term. Players can also earn points, but only to increase player gratification.

For von Ahn, the ESP Game is more than a diversion — it could be a viable way to “train” computers to identify images. “Computers can learn by example, and with online images, computers would have an enormous base of examples.”

Image-savvy computers could be useful in medicine and security, to name just a few potential fields; von Ahn described the possibility of computers that could identify medical anomalies in X-rays or military bunkers in aerial photographs, all without the aid of a human.

Google uses their Image Labeler for a slightly different purpose. With a constant influx of online images, a game that encourages players to label as many pictures as possible is a fun, productive, and inexpensive way to improve Google Image Search.

As to how Google incorporates their images into their database, von Ahn is unsure. “They do as they please with it,” he said. He had received multiple offers prior to licensing the game to Google, and “it took a couple of months” to decide to go through with the transaction.

When asked why he chose Google, von Ahn laughed and said, “I don’t think I want to say why.” Meanwhile, he still holds the rights to develop the ESP Game for research purposes.

He also holds the rights to several other online games he’s developed. These games include a similar image-labeling game called Peekaboom; a word-identification game called Verbosity; and Phetch, an image-retrieval game that could one day help blind Internet users identify images on the Web.

“This has been the past three or four years of my life,” said von Ahn. In addition to his work with human computation, von Ahn teaches the course Great Theoretical Ideas in Computer Science (15-251).

Although von Ahn declined to comment on the details of the transaction with Google, his office in Wean Hall did show evidence of some compensation. The only place to sit, besides at the desk, is in an enormous, primary-colored Google beanbag chair.