Cyclic deep representation-based domain adaptation for cross-subject motor imagery classification

Deep representation learning has attracted great attention for brain-computer interfaces (BCIs) based neural rehabilitation engineering, especially for the motor imagery electroencephalogram (MI-EEG) signals. Recently researchers have explored numerous deep representation models to decode MI-EEG signals with various structures, however they suffered from the variability across recording subjects and the scarcity of samples. To solve these issues, domain adaptation models have been proposed to mitigate existed subjects’ samples to decode new subject’s sample by learning subject-invariant deep representations. However, existed models neglected temporal-varying and spatially-coupled characteristics of MI-EEG signals during domain adaptation, resulting performance deterioration for cross-subject classification. To improve decoding performance, we propose a novel domain adaptation model, referred to C yclic D eep R epresentation-based D omain A daptation (CDRDA), to simultaneously transfer deep representations from source domain to target domain, as well as target domain to source domain. Specifically, our CDRDA model learns a joint optimization that weighted dual adversarial losses, cyclic losses, and domain-specific losses to improve classification performance together. Empirical experiments on two benchmark MI-EEG datasets have revealed the feasibility and effectiveness of the CDRDA model with accuracy, Cohen’s kappa, and macro average F1-score. Results analyses and ablation studies have also verified the superiority of the CDRDA model for building online MI-BCIs.