Human Computation

A friend of mine recently pointed me to this video by a guy called Luis von Ahn. It looks like his work has gotten a lot of coverage in the blogosphere. I would highly recommend the talk. Though provoking to say the very least. You don’t really need a technical background to be able to understand what he is talking about.

The gist of von Ahn’s work lies in being able to cleverly leverage what he calls wasted human cycles. People spend gazillions of human hours playing games that have no obvious use. So why not devise games that end up solving real world problems? That’s what von Ahn has done – ESPgame and Peekaboom. What ESPgame does is it pairs you up with an anonymous partner. It supplies you with a series of images that it asks you to label. You win points if you and your partner agree upon the same labels. So what’s the deal with leveraging human cycles here? In the process of playing players inadvertently end up improving the quality of image search results. The vast majority of images on the Internet are unlabeled and state of the art search engines use (not so clever) heuristics such as the name of the image file and words that can be found in the proximity of the image. Surprise surprise.. in the aftermath of von Ahn’s talk at Google, the folks in Mountain View have licensed his game to improve the quality of Google image search.

While the notion of human computation (as applied to image search) is interesting to say the very least, I suspect that it will not work well in isolation. Why? Because it will be ineffectual while catering to the long tail. Given the picture of a person, most players are likely to use the labels “man” or “woman” since they have no clue as to who that person is. For example, a label of “man” to describe a picture of the President of Uruguay is unhelpful. Here’s another example.. Since there is very little by means of visual information to distinguish this picture of Lake Tahoe from this one of Crater Lake, most players (who have not been to either place) are likely to use tags like “scenery” and “lake” that while being accurate in describing the picture are way too generic when it comes to actually making a qualitative difference to search experience.

On the flipside, the notion of human computation did not seem totally unfamiliar. That’s what del.icio.us users have been doing for several years now. They end up improving search results by tagging/book marking their URLs. I suspect that del.icio.us results are likely more accurate than those generated by the ESPGame. The reason being folks using del.icio.us have more information to describe a page that a URL refers to (having actually read it) as opposed to people who are presented with a random image of a person whom they have never seen before.