This is a black-box learning challenge: Competitors train a classifier on a dataset that is not human readable, without knowledge of what the data consists of. They are scored based on classification accuracy on a private test set. This challenge is designed to reduce the usefulness of having a human researcher working in the loop with the training algorithm.
We are also providing a dataset of approx. 130,000 unsupervised examples that contestants can use to improve their models. The unsupervised data is a CSV file in the same format as the private test set (i.e. without the labels). The extra data comes from a distribution that is very similar to the training/test set distribution.
We provide example code for this contest as part of the pylearn2 package at
[https://github.com/lisa-lab/pylearn2](https://github.com/lisa-lab/pylearn2)
For this contest, look at the pylearn2/scripts/icml_2013_wrepl/black_box directory.