Graph Learning With Co-Teaching for EEG-Based Motor Imagery Recognition

Previous studies have explored the use of deep neural networks for electroencephalography (EEG)-based motor imagery (MI) recognition, but most of the models focus on the recognition performance achieved for a single subject and are challenging to transfer due to individual differences and low signal-to-noise-ratio of EEG signals. To date, few studies have paid attention to the balance between generalizability and personalization across subjects. To this end, we propose a co-teaching graph learning method for cross-subject EEG-based MI recognition. First, A novel graph learning approach is designed to improve feature extraction from a typical graph structure containing raw EEG signals. Second, two graph learning models are constructed to filter noisy data by using a co-teaching training strategy, preventing overfitting on noisy samples obtained from different subjects. The proposed model shows a 5.4% and 3.2% increase in accuracy of single- and multisubject four-class MI recognition tasks compared to the previous best method, respectively. Experimental results also demonstrate that it is easy to derive a model that can represent generic knowledge of multiple MI subjects and can be fine-tuned efficiently for new subjects.