Robustness of the Common Spatial Patterns algorithm in the BCI-pipeline

When we want to use brain-computer interfaces (BCI) as an input modality for gaming, a short setup procedure is necessary. Therefore a user model has to be learned using small training sets. The common spatial patterns (CSP) algorithm is often used in BCI. In this work we investigate how the CSP algorithm generalizes when using small training sets, how the performance changes over time, and how well CSP generalizes over persons. Our results indicate that the CSP algorithm severely overfits on small training sets. The CSP algorithm often selects a small number of spatial filters that generalize poorly, which can have in impact on the classification performance. The generalization performance does not degrade over time, which is promising, but the signal does not seem to be stationary. In its current form, the CSP generalizes poorly over persons.