Multi-view Contrastive Learning for Unsupervised Domain Adaptation in Brain-Computer Interfaces
Abstract
Electroencephalography (EEG) has been widely used to record electromagnetic fields for motor imagery (MI)-based brain-computer interfaces (BCIs). However, collecting MI signals is often time-consuming and challenging to classify due to the inter-subject variability of EEG signals. To address these issues, we propose a novel framework MACNet, which stands for Multi-view Adversarial Contrastive Network. MACNet employs a contrastive learning approach to learn spatial and temporal features in two views, using Riemannian and Euclidean encoders. By jointly extracting underlying features and learning domain-invariant representations in both source and target features, MACNet improves the alignment and accuracy. In addition, we propose a domain mixup for the BCI field at the signal and embedding levels, to improve domain alignment. We evaluate MACNet on two public datasets and demonstrate that it outperforms all previous methods in inter-subject transfer learning. Specifically, MACNet achieves 83.79\% accuracy for the BCI Competition IV dataset and 80.00\% accuracy for the OpenBMI dataset.