# MI-023: Transfer Learning for Motor Imagery Based Brain-Computer Interfaces: A Complete Pipeline > Generated from a local paper card. Do not edit this page directly; edit the source card and rerun `vp run docs:generate`. > Internal wiki boundary: paper-card evidence and user-provided PDF downloads are available for private reading. Extracted full-paper text is not published. > Source: `library/paper_cards/MI-023.md` ## Paper Access * Internal PDF: Download Paper * DOI / official page: [needs confirmation](https://arxiv.org/abs/2007.03746) * Open-access page: [Open access page](https://arxiv.org/abs/2007.03746) * Deployment boundary: these PDF links are intended only for a private/protected internal wiki. ## MI-023: Transfer Learning for Motor Imagery Based Brain-Computer Interfaces: A Complete Pipeline ## Metadata * ID: MI-023 * Title: Transfer Learning for Motor Imagery Based Brain-Computer Interfaces: A Complete Pipeline * Year: 2021 * DOI / URL: https://arxiv.org/abs/2007.03746 * Local PDF: see Paper Access section above * Text artifact: local-only path withheld from docs site * Review status: `extracted` ## Study Type * Track: MI / transfer learning * Task: complete transfer-learning pipeline for MI BCI * Participants or dataset: two MI datasets; exact dataset names and counts need extraction * Device/electrode setup: EEG MI datasets with standard channel structures * Protocol/task: calibration reduction for MI BCI ## Methods * Signal processing or analysis: Euclidean alignment, Riemannian alignment, CSP, feature selection, classifiers, and transfer blocks * Training/calibration: explicitly targets reducing calibration effort for new subjects * Online/offline: offline calibration experiments ## Key Results * The abstract states that offline experiments on two MI datasets verified improved performance. * The paper argues that transfer learning should be considered across spatial filtering, feature engineering, and classification, not only one component. ## Limitations * arXiv/preprint version; final publication metadata needs confirmation. * Exact quantitative results require table extraction. * Offline calibration does not prove online no-control safety. ## Relevance To Current Review * Strong recent source for the MI calibration-reduction trend. * Useful to plan Exp2 if the project wants to compare subject-specific, cross-subject, and transfer-adapted decoders. ## Evidence Status | Claim | Status | Evidence Note | | --- | --- | --- | | MI transfer learning can be applied across multiple BCI pipeline blocks. | verified | Abstract explicitly names spatial filtering, feature engineering, and classification. | | Calibration effort is a central problem for MI BCI. | verified | Abstract frames TL around reducing calibration for a new subject. | | Transfer learning removes the need for online validation. | needs confirmation | Online closed-loop and no-control evidence remain required. | ## Open Questions * Which transfer block should be implemented first: alignment, feature adaptation, or classifier adaptation? * Does the project have enough subjects to evaluate transfer meaningfully?