MixDual-Tuning: Improved Fine-Tuning for Cross-Subject Few-Shot Motor Imagery Classification

Abstract

Motor imagery (MI) brain-computer interfaces (BCIs) face challenges posed by individual differences, and models trained on existing subjects are difficult to apply directly to target subjects. Although transfer-learning-based approaches can help alleviate this problem, they require a large amount of target data for model fine-tuning, which leads to a heavy data collection burden and causes mental fatigue in subjects. Recently, several few-shot learning-based approaches have been applied to MI-BCI, achieving promising performance with a small amount of data on target subjects. However, most existing techniques ignore the overfitting and domain bias issues associated with limited data. To address these challenges, we propose the MixDual-Tuning method, a novel fine-tuning method that combines data augmentation with an improved dual-component loss function. In detail, synthetic data are first generated through augmentation and then combined with target domain data to increase data volume and diversity, reducing overfitting. Moreover, a dual-component loss function is applied to encourage domain-invariant feature learning, by combining cross-entropy loss for classification and optimized MMD loss for domain alignment. We also introduce and enhance a flooding regularization technique to prevent overfitting by stabilizing the training loss. We evaluate the effectiveness of MixDual-Tuning on three publicly available MI-BCI datasets, including competition IV-2a, IV-2b, and our dataset. Extensive experiments demonstrate that MixDual-Tuning consistently surpasses both baseline models and recent few-shot learning approaches, verifying the effectiveness of the proposed method.