How to Spot a Psychopath

I only yesterday got around to watching Luis von Ahn's excellent Google TechTalk from last year on Human Computation.

It's very interesting, though like totally the outside scoop, man, for people who follow the world of human-versus-computer data analysis.

I was pointed to it by comments on the Coding Horror post on whether Amazon's Mechanical Turk is a failure. Von Ahn's insight is that you don't have to pay people to do many seemingly tedious tasks which humans can do better than computers. If you can make a game of it, they'll do it for free.

The comments also, of course, point to von Ahn's The ESP Game, a perfect example of the theory in action in which anonymous pairs of people play a timed game of "Snap" in their attempts to type the same word when shown the same image, as a result creating a database of labels for those images.

Later on, there's a mention of Google Image Labeler, which is an exact (licensed) copy of The ESP Game. The difference is that Google Image Labeler appears to be working on the actual Google Images database. It's therefore doing real image-labelling work, as well as providing the entertainment that can only be gained when you boggle at your partner's apparent complete inability to recognise a picture of a shoe.

The ESP Game is more of a research tool, so it only works on a more controllable 30,000 image database. That database has to be about as well-labelled as it's ever going to get, by now.

(My own lame take on this idea is this piece. I don't think it'll be long before we see a game like Left 4 Dead or Natural Selection in which paying customers can play either side, but freeloaders can only be zombies/aliens/kobolds.)

While I'm linking to cool new information processing ideas that most of you dorks have probably already seen, allow me to highly recommend Scene Completion Using Millions of Photographs. The 11Mb PDF is well worth downloading.