The goal of this challenge was to detect error related potential recorded during a p300 spelling task. The classification must be done across subjects, i.e. training and test sets were composed by different subjects. This is a hard task due to the high inter-subjects variability of EEG. However, my Riemannian Geometry framework has been proven very powerful for dealing with this problem.
The main difficulties of this challenge were to deal with the relatively high number of electrodes and to avoid overfitting. We overcome the first issue by using a channel selection algorithm, and the second by using a bagging procedure and an appropriate cross-validation methodology.