Incremental versus non-incremental learning in adaptive common spatial patterns

Abstract

Extracting reliable and discriminative features remains a critical challenge in the development of brain computer interface (BCI) techniques. Common spatial patterns (CSP) is frequently used for spatial filtering and feature extraction in electroencephalography (EEG)-based BCI. It performs a supervised and subject-specific learning of EEG data acquired in two different task conditions. Incremental learning has been used in CSP to adapt to a target subject's data by including classified data in training data and re-estimating spatial filters. In practical circumstances where no user feedback is instantly available to provide true class labels of target trials, misclassified EEG trials will be added to the training data of a wrong class, and potentially influence the training of spatial filters and feature extraction. In this study, incremental and non-incremental learning were investigated based on a recently developed adaptive CSP (ACSP) method using multi-subject EEG data. Their performances were compared in terms of intra- and inter-subject classification performances. Experimental results indicate that the non-incremental learning is a better option when true class labels of target data are not provided.