recent updates

help


User login

Google Image Labeler -- Game Theory and Information

By Zach Whalen – Mon, 2006 – 09 – 04 16:42

Google has introduced an interesting and highly addictive game as part of its image search component. Google Image Labeler works by pairing you with another random player and throwing images at you. Your collective goal is to type in text labels for the image until you and your partner come up with an identical label. Having done so, you get 100 points and move on the next one. You have 90 seconds... Go.

In case this seems familiar, it's basically the inverse of "Guess the Google".

Its fast pace keeps it addictive, and the running leader board encourages you to keep going. With a good partner, I've found it pretty easy to come within striking range of the current top 10.

There's some speculation at Webmasterworld that Google is using this to sharpen its image searching algorithm or to build a database of tagged images a la Flickr, but if that's the case, I don't know how useful these tags will be. It's very hard to tag portraits or simple pictures of an individual (for example), so you start throwing out every noun you can associate with it. You end up with something like "hair" or "shirt" or "person." This strategy can also be exploited directly, as indicated by the nickname of the points leader when I logged in yesterday: "just write t**s."

I hope someone who really understands game theory can unpack this, but the Google Image Labeler game seems to me an interesting case of a non-zero-sum game. Something like an iterated prisoner's dilemma where cooperation also benefits the police (Google). The only way to gain unfair advantage over the "police" in this scenario is through a pre-arranged contract (just write t**s). So for this project to be successful from Google's point of view, they have to have calculated the extent to which successful cooperation depends on specificity in labeling, but a pre-arranged contract breaks that calculation.

I'm really not sure if that helps explain why this game is so interesting (and addictive), but I would appreciate it if anyone can correct or amend my application of game theory here. I just think that since the structure of this Image Labeler game is so similar to the scenarios used in explaining game theory, there must be a way to talk about it in those terms.

I also think this game might qualify as a blend between serious and casual gaming, but this post is getting long.

Any thoughts?

a.k.a. The ESP Game

Submitted by Thlayli (not verified) – Wed, 2006 – 09 – 06 18:20

In case it seems really familiar. It's a perfect rip-off of The ESP Game. If they got permission to use the format, they sure didn't credit CMU.

Wow

Submitted by Zach Whalen – Thu, 2006 – 09 – 07 08:36

Yeah, that is remarkably similar, even down to the phrasing of the messages. What I'm not clear on, though, is that the ESP game also claims to help improve image metadata, but I don't see what they're doing with the data they gather.

The Google version does make a few things easier by using Ajax for interaction, rather than Java, but the two games do some conceptually identical. I do think, though, that if we're talking about a prisoner's dilemma in which the third party is the police, Google seems more like the "police" in its quest to assimilate all of the world's information than CMU does.

Thanks for the link.

Great minds use the same labels

Submitted by Jeremy (not verified) – Sun, 2006 – 09 – 10 19:39

I had a very similar reaction, commenting a comparison to the Amazon mechanical Turk.

I wonder if it is eerie or obvious that we both immediately thought of the iterated prisoner's dilemma? Perhaps only because it leads to great examples of collusion. Reactions:

Going by title, "Image Labeler" this is being presented as a collective intelligence public works project, like Amazon Mechanical Turk. That may be why the side-effects of competitive play weren't fully anticipated - the attitude may have been that to some extent all play would be collectivist in outlook. Perhaps a hard lesson will be learned when they datamine the results, or perhaps they have found that it is trivial to screen cheaters - although it doesn't appear so given that the rankings boards have been full of cheaters.

Also, regarding the comment on the ESP game, Luis von Ahn has confirmed it was Google licensed for Image Labeler purposes.

Google Image Labeler Cheaters

Submitted by Wraith (not verified) – Tue, 2007 – 05 – 08 15:24

You are correct that those in the top ranking are cheaters. They seem to get some sort of jolt out of this "power trip" having their names on the big board.

The main culprit is an ancient professor, Frank D, from Rutgers University. He must be tenured because he spends and inordinate amount of time pouring over results from that game.

Rutgers must have a lot of money to waste if their professors have nothing to do other than spend a lot of their time playing online games.

my professor

Submitted by MP (not verified) – Wed, 2008 – 03 – 19 00:40

yeah he was my professor - he actually does a lot of research and is very cultured

I suspect that he doesn't sleep much.