A study on the application of contrastive learning in the brain-computer interface of motor imagery

Abstract

In brain-computer interface (BCI) systems, acquiring EEG signals is relatively easy, but labeling EEG signals is challenging. Supervised learning is widely used in existing BCI research, but its generalization ability is poor due to the influence of labeled data, which is not conducive to the promotion and application of BCI. In the fields of computer vision and natural language processing, self-supervised learning is becoming a popular research direction that can use unlabeled data for learning. In this study, we propose a self-supervised contrastive learning method for identifying motor imagery states by performing different data transformation operations on EEG data to generate positive and negative sample pairs. This allows the pre-trained network to learn general feature representations from the signals, enabling it to converge faster and achieve better classification performance when performing downstream tasks. The network's performance is significantly related to the amount of unlabeled data used in the self-supervised learning process. Through the evaluation of the collected data, we achieved an accuracy of 82.56% in classifying motor imagery states with 40% amount of labeled data, demonstrating the practical value of our method.