Domain generalized feature embedded learning for calibration-free event-related potentials recognition.

Abstract

Event-related potentials (ERPs) play an important role for building EEG-based brain computer interfaces (BCIs). However, due to the complex and varied spatio-temporal characteristics of ERPs across subjects, data distribution across subjects a very important issue to be solved for constructing calibration-free BCIs. To achieve calibration-free ERPs recognition, we propose a D omain G eneralized F eature E mbedded L earning (DGFEL) method. First, we align the ERPs of each existed subject based on covariance centroids. Then, we enhanced the aligned samples based on xDAWN filter and extract spatio-temporal features. Finally, the spatio-temporal features are further generalized by the decomposed adversarial loss, and we construct a neural network embedding backbone to implement features generalization across subjects. The proposed method has been systematically validated on two benchmark EEG-based ERP datasets, and its classification performance surpasses several state-of-the-art methods as well as deep learning models. Moreover, it effectively captures robust features from existed source subjects, and can be generalized to new subjects without accessing target ERP samples. Our method therefore provides a novel selection to construct calibration-free ERP-BCIs.