Self-supervised Contrastive Few-Shot Learning for Motor Imagery Brain-Computer Interfaces.

Motor imagery brain-computer interfaces (BCIs) are integral to the field of intelligent rehabilitation. However, the collection of motor imagery data is time-consuming and expensive, while traditional deep learning models often face the risk of overfitting when handling limited data samples. In this paper, we propose a motor imagery recognition network that merges a self-supervised contrastive framework with few-shot learning. The proposed model is first trained on the dataset without labels using self-supervised contrastive to obtain the optimal model. The model parameters obtained are then used as initialization parameters for the prototypical network, which improves in few-shot learning. The results show that on the BCICIV2A and BCICIV2B datasets, the proposed approach achieves accuracies of 69.12% and 80.27%, respectively, representing improvements of 6.79% and 4.14% compared to the baseline models. This research achievement holds significant implications for advancing the application of motor imagery BCIs in the domain of intelligent rehabilitation.