The idea is to ask you to classify as signal (Higgs decay to tau lepton pairs, if you really want to know!) or background (anything that looks similar to it but involves no Higgs boson) a set of 550,000 events, for each of which ATLAS gives you 30 kinematical quantities measured in the detector (it is a simulation, but it's a pretty good approximation of reality).
How are you supposed to decide whether an event is signal or background ? You can do that by using a training set of 250,000 events, mixed signal and background, which are similar to the ones you need to classify, but which also contain the label "s" or "b". So if you can write a program that uses the differences in the 30 kinematical distributions for signal and background events, and learns to tell them apart, then you can classify the 550,000 test events and win a prize !
Because it is of real money that we are talking. The first prize is US$7000, the second is US$4000, and the third is US$2000. And you do not even need to hurry: there are about four months to go, and you can enlist until the beginning of September.
Most participants (160 so far, and counting) are not waiting idly. The nice thing about the competition is that you can submit an answer, and it will result in a score. The score is not the final score that your answer would eventually get, as it is computed by ignoring 82% of the sample; but it is a good approximation of that. So you can submit an answer, and see how well you do with respect to your competitors. It is a rush! You can submit up to 5 sets per day.
I find this a very nice way to understand what multivariate algorithms are better performing at the classification task which the problem constitutes. It is a quite typical setup for high-energy physics, and indeed ATLAS plans to learn from the code of the winners (which will have to be submitted at the end of the process).
If you want to participate, just enlist at the kaggle site. Or just have a peek at the leaderboard, which is in continuous evolution. It is fun !