Background/Objectives: Motor imagery (MI) EEG-based brain–computer interfaces (BCIs) are promising for neurorehabilitation, but practical use is often hindered by time-consuming per-user calibration and performance instability across sessions/users. Methods: To mitigate this issue, we aim to improve subject-dependent MI classification by leveraging labeled training data from other subjects within the same dataset via transfer learning. We propose Maximizing Single-Feature Separability (MSFS), a lightweight plug-in regularization applied during target–subject fine-tuning. MSFS operates on the network feature layer and constructs batch-wise target positions by maximizing a silhouette-based separability criterion for each feature dimension. The target position computation is implemented in a fully vectorized GPU-friendly manner. Results: We evaluate MSFS on BCI Competition IV-2a and IV-2b datasets using three representative backbone networks (EEGNet, ShallowConvNet, ATCNet). MSFS consistently improves standard transfer learning across both datasets and all backbones. When compared against representative transfer learning algorithms from the literature, MSFS remains competitive against the literature baselines. Ablation analysis confirms the effectiveness of each algorithm component. Few-shot experiments further indicate that MSFS is still beneficial when the target subject provides limited labeled data. Conclusions: MSFS provides a within-dataset transfer learning enhancement for MI EEG decoding, improving target–subject accuracy under limited calibration data without relying on external datasets, and can be readily integrated into common deep MI classification pipelines.
Xu et al. (Sat,) studied this question.