GINA is digit recognition database

The task of GINA is handwritten digit recognition. For the “agnostic learning track” we chose the problem of separating two-digit odd numbers from two-digit even numbers. Only the unit digit is informative for that task, therefore at least ½ of the features are distracters. Additionally, the pixels that are almost always blank were removed and the pixel order was randomized to hide the feature identity. For the “prior knowledge track”, only the informative digit is provided in the original pixel map representation. This is a two class classification problem with sparse continuous input variables, in which each class is composed of several clusters. It is a problems with heterogeneous classes.