EEG-Based Motor Imagery Classification with Deep Adversarial Learning

Abstract

Brain-computer interface (BCI) provides a direct communication pathway from the human brain to computers. In consideration of electroencephalogram (EEG) data are greatly affected by the individual differences, it is hard to acquire general models applicable across subjects. For this reason, we propose a domain adversarial neural network (DANN) that can efficiently extract and classify domain-invariant features across domains. DANN consists of three parts: feature extraction, label prediction, and domain discrimination. During training, labeled data of the source domain and unlabeled data of the target domain are used as inputs. The part of feature extraction learns discriminative features of EEG signals through deep networks. The features of the source domain data are then fed into the label prediction part for learning to distinguish between different task classes, while the part of domain discrimination reduces inter-domain differences by learning domain-invariant features across domains. By optimizing both the part of label prediction and the domain discrimination, the model learns task classification features that are domain-invariant. DANN approach is validated on several BCI competition datasets, indicating its advantages in cross-subject motor imagery classification.