Classifying regularized sensor covariance matrices: An alternative to CSP

Abstract

Common spatial patterns ( CSP) is a commonly used technique for classifying imagined movement type brain-computer interface ( BCI) datasets. It has been very successful with many extensions and improvements on the basic technique. However, a drawback of CSP is that the signal processing pipeline contains two supervised learning stages: the first in which class-relevant spatial filters are learned and a second in which a classifier is used to classify the filtered variances. This may lead to potential overfitting issues, which are generally avoided by limiting CSP to only a few filters. This work argues for an alternative approach where only a single supervised learning stage is needed. The key step in this approach is to use whitened spatial covariance matrices as features and to use a linear classifier to simultaneously learn the spatial filters and the classifier weights. Unfortunately, this approach can also lead to overfitting problems. We show how these problems can be addressed by appropriately regularizing the whitening computation. Ridge regularized covariance classification outperforms whitened spatial covariance, CSP, and two regularized CSP classification methods on an imagined movement dataset indicating the usefulness of this regularization method for BCI. Trace norm regularization can help with the interpretability of the results.