Multitask Learning for Common Spatial Patterns

Common spatial pattern (CSP) is a spatial filtering algorithm which can be used in motor imagery-based braincomputer interface (BCI) for extracting spatial feature of multivariate signals. But CSP algorithm has inherent drawback, which is that the estimation of the covariance matrices is sensitive to noise, and when few data are available, CSP algorithm is very likely to overfit while it is very time consuming to collect large amount of training data from each task. In this paper, we propose a multitask learning method to extract discriminative subspace shared between subjects and regularize CSP away from the orthogonal complement of this subspace. We compared our method with the standard CSP algorithm on three publicly available datasets of BCI competitions, and an significant performance gain was observed, which therefore demonstrated the advantage of the proposed method.