Multi-source Dictionary Transfer Learning for few-shot motor imagery EEG classification

Motor imagery electroencephalography (MI-EEG) is a critical brain-computer interface paradigm for direct control of external devices. However, cross-subject MI-EEG classification faces two major challenges: (1) the complex dynamic nature and low signal-to-noise ratio of MI-EEG makes it difficult to extract high-level features that have enough class-discrimination; (2) the scarcity of labeled target-subject data severely limits model generalizability. To address the problems, this paper proposes a novel method called multi-source dictionary transfer learning (MSDTL). Through Fisher-embedded dictionary atoms and coefficients, MSDTL enhances intraclass compactness while maximizing interclass separation. Simultaneously, maximum mean discrepancy-based adaptation aligns marginal and conditional distributions across domains. Extensive experiments on four MI-EEG datasets demonstrate MSDTL’s superiority. It achieves 84.36% average accuracy on the BCI Competition IV Dataset I under several few-shot settings, outperforming state-of-the-art methods by 0.2%-4.06%. While current implementation focuses on offline analysis, future work will extend to real-time and semi-supervised scenarios with unlabeled data.