Online feature selection for Brain Computer Interfaces

Abstract

Online adaptation of Brain Computer Interfaces allows for arduous training periods to be circumvented. To do this we must adapt a classifier to a new session, or better yet, a new subject. We initially outline a procedure to perform online adaptation of both the classifier's weights and the feature selection and confirm its use in session to session transfer. We found that retraining both feature selection and the classifier resulted in an average improvement of 5% over simply retraining the classifier, and as high as 10%. To avoid a retraining phase the online adaptation must be performed without labeled data. We propose and compare several methods to adapt the feature selection on unlabeled data, making use of both semi-supervised learning and interactive error potentials. From this we determined that performing a weighted feature selection performed the best, and the proposed novel approach of combining semi-supervised learning and interactive error potentials outperformed performing each individually. To improve the subject to subject adaptation when a database of previous subjects is available, we investigated using Weighted Majority Voting to weight the classifier towards subjects in that database that are useful for the new subject. We found this approach to outperform pooling all data.