Comprehensive deep representation-based Wasserstein domain discriminator for cross-subject motor imagery EEG classification.
TL;DR
A novel deep domain adaptation model, referred to as Comprehensive Deep Representation based Wasserstein Domain Discriminator (CDR-WDD), to simultaneously represent comprehensive deep features and adversarial learned based on the domain discriminator with stable learning gradient is proposed.
Abstract
Motor imagery electroencephalogram (MI-EEG) signals have attracted great attention for assistant rehabilitation in biomedical engineering. However, due to recording environment, device, and subject variabilities, original deep neural network models are suffered from data distribution discrepancies across subjects. Although recently deep domain adaptation models endeavored to solve the distribution discrepancies, they suffer from the simple feature representation and gradient vanish during domain-invariant feature learning. To this end, we propose a novel deep domain adaptation model, referred to as Comprehensive Deep Representation based Wasserstein Domain Discriminator (CDR-WDD), to simultaneously represent comprehensive deep features and adversarial learned based on the domain discriminator with stable learning gradient. Specifically, our CDR-WDD model involves a jointly optimization that weighted the loss of feature representing and adversarial learning, which improves the performance of cross-subject MI-EEG classification. Empirical studies on four benchmark MI-EEG datasets have revealed the superiority performance of the proposed model compared with state-of-the-art deep domain adaptation models, which achieves average accuracies and Cohen's kappa value of 84.34, 89.59, 73.32, 87.75 and 0.790, 0.802, 0.466, 0.755 for BCIIV-2a, BCIIV-2b, OpenBMI, and WCCI datasets, respectively. Ablation studies have been conducted to show the effectiveness of CDR and WDD modules. Our CDR-WDD model provides a novel option to build brain-computer interfaces.