A Debiased Domain Adaptation Framework with Minimum Class Confusion for Motor Imagery Decoding

Cunhang Fan,Zhen Chen,Xun Song,Xinhui Li,Jun Xue,Ping Li,Zhao Lv
DOI: https://doi.org/10.1109/ijcnn60899.2024.10649993
2024-01-01
Abstract:Recently, motor imagery decoding technology based on electroencephalogram (EEG) signals has made significant progress. However, there are still challenges in adapting to new sessions, mainly due to changes in data distribution between different sessions and confusion problems caused by similar oscillation patterns in various categories of EEG signals. To address these problems, this paper proposes a debiased domain adaptation framework with minimum class confusion to learn unbiased representations in motor imagery tasks. Specially, unlike the feature alignment and adversarial training methods, we explore the class predictions for domain adaptation, applying the minimum class confusion loss criterion in the target domain to reduce inter-class confusion. This approach aims to alleviate the bias issues inherent in classifiers trained on the source domain for making predictions in the target domain, achieving class-level alignment. Consequently, it enhances the model’s ability to adapt to the data distribution of the target domain. Experimental results on two public EEG datasets (BCI Competition IV datasets IIa and IIb) show that the method for cross-session decoding is significantly improved compared to the baseline, with average classification accuracy reaching 81.01% and 82.92%, respectively.
What problem does this paper attempt to address?