Deep Neural Network with Joint Distribution Matching for Cross-Subject Motor Imagery Brain-Computer Interfaces

Figure 1

Illustration for joint distribution adaptation. DS: source domain, red solid circle; DT: target domain, black dashed circle. +: centroid of positive class; -: centroid of negative class. Red hollow circle: centroid of source data; black solid circle: centroid of target data. fS: discriminative line for source data; fT: discriminative line for target data. Marginal domain adaptation (MDA) utilizing MMD makes the centroid of source data (red hollow circle) and that of the target data (solid black circle) closer. Joint distribution adaptation aligns source and target data according to the class they belong to. It makes the discriminative lines of source and target most similar, and the classifier trained with source data will be transferred to target most effectively.