Abstract
• A gated parallel feature fusion multi-task learning model is proposed to solve data distribution discrepancy in Motor Imagery EEG classification.
• The proposed model strikes the proper balance between task-specific learning and shared knowledge learning.
• Extensive experimental results show our superior performance over representative models under the cross-subject evaluation.
• The source code of our proposed model can be found in https://github.com/Henrywang621/GPFMTL-for-MI-EEG-classification.
Motor imagery (MI) EEG-based brain-computer interfaces (BCIs) allow users to control external devices with their thoughts, which has been widely used in both medical and commercial fields. Recent advances in deep learning (DL) have greatly facilitated the rapid development of MI EEG-based BCIs. However, an open issue (i.e., data distribution discrepancy) hinders the further improvement of the performance in this field, especially for cross-subject classification scenarios. Although several studies have developed multi-task learning (MTL) to solve this issue, their shared feature extractor limits models’ ability to learn task-specific features. In this paper, we proposed a gated parallel feature fusion MTL model named GPFMTL for MI-EEG classification to mitigate data distribution discrepancy—further improving classification performance. Unlike the previous works, our design strikes the proper balance between task-specific learning and shared knowledge learning. This is achieved by developing tailored subnets for different related tasks to better extract task-specific features and adopting gate control modules to effectively control information sharing between different subnets. The extensive experiments show the proposed model can significantly outperform eleven state-of-the-art DL-based models in some cases on two public datasets (p-value < 0.05) under the cross-subject evaluation, achieving improvements of approximately 2.1 % on the PhysioNet dataset and 2.6 % on the BCI IV 2a dataset, respectively. Besides, a series of ablation studies are conducted to further explore the effectiveness of the proposed models. Our findings could help future studies better develop MTL in this field.