Classification of Structured EEG Tensors Using Nuclear Norm Regularization: Improving P300 Classification

Abstract

Choosing an appropriate approach for single-trial EEG classification is a key factor in brain computer interfaces (BCIs). Here we consider an auditory oddball paradigm, recorded in normal indoor and walking outdoor conditions. The signal of interest, namely the P300 component of the event related potential (ERP), unlike noise, is a structured signal in the multidimensional space spanned by channels, time and frequency or possibly other types of features. Therefore, we apply spectral regularization using nuclear norm on a tensorial representation of the EEG data. Due to the a-priori structural information conveyed by the nuclear norm penalty, we expect an improved performance compared to traditional approaches, especially under noisy conditions and in case of small sample sizes.