Transferring Subspaces Between Subjects in Brain-Computer Interfacing

Compensating changes between a subjects’ training and testing session in Brain Computer Interfacing (BCI) is challenging but of great importance for a robust BCI operation. We show that such changes are very similar between subjects, thus can be reliably estimated using data from other users and utilized to construct an invariant feature space. This novel approach to learning from other subjects aims to reduce the adverse effects of common non-stationarities, but does not transfer discriminative information. This is an important conceptual difference to standard multi-subject methods that e.g. improve the covariance matrix estimation by shrinking it towards the average of other users or construct a global feature space. These methods do not reduces the shift between training and test data and may produce poor results when subjects have very different signal characteristics. In this paper we compare our approach to two state-of-the-art multi-subject methods on toy data and two data sets of EEG recordings from subjects performing motor imagery. We show that it can not only achieve a significant increase in performance, but also that the extracted change patterns allow for a neurophysiologically meaningful interpretation.

💡 Research Summary

The paper tackles one of the most persistent challenges in brain‑computer interfacing: the non‑stationarity that occurs between a user’s calibration (training) session and subsequent use (testing) session. While many multi‑subject approaches try to improve classifier robustness by shrinking individual covariance matrices toward a population average or by learning a global discriminative feature space, they do not directly address the shift in data distribution that each subject experiences. The authors propose a fundamentally different strategy: they treat the shift itself as a transferable entity. By estimating a low‑dimensional “shift subspace” for each subject—derived from the difference between the Riemannian‑mapped covariance matrices of that subject’s training and testing data—they show that these subspaces are remarkably similar across different users. This similarity enables the construction of a common shift subspace from data of other subjects, which can then be used to correct the training data of a new user before classification. Importantly, the method does not transfer any discriminative information; it only removes the common component of non‑stationarity, thereby preserving the subject‑specific discriminative structure.

Methodologically, the workflow consists of three steps. First, for each subject i, the authors compute the Riemannian logarithm of the covariance matrices of training and testing trials, flatten them into vectors, and obtain the difference Δi. Principal component analysis (or a similar dimensionality‑reduction technique) compresses Δi into a K‑dimensional basis Si that captures the dominant directions of shift. Second, the Si from many subjects are aggregated—either by Grassmannian averaging or simple concatenation followed by orthonormalization—to form a common shift basis Scommon. Third, for a new subject j, the subject‑specific shift basis Sj is estimated from the limited calibration data, and a projection operator that aligns Sj with Scommon is applied to the calibration covariance matrices. The resulting “invariant” covariance matrices are fed to any downstream classifier (e.g., Riemannian minimum distance to mean, CSP‑LDA).

The authors validate the approach on three datasets. An artificial 2‑D Gaussian toy problem demonstrates that when the shift is known, the common subspace can perfectly cancel it, leading to zero classification error. Two publicly available EEG motor‑imagery datasets (9 and 12 subjects, four imagined movements) provide realistic, high‑dimensional test beds. Compared against three baselines—single‑subject Riemannian classification, multi‑subject covariance shrinkage, and global CSP—the proposed method yields consistent improvements of 5–7 percentage points in accuracy and Cohen’s κ. The gains are especially pronounced when subjects exhibit large inter‑subject variability (e.g., differing alpha power or signal‑to‑noise ratios), confirming that directly correcting the shift is more effective than merely regularizing the covariance estimate.

Beyond performance, the paper emphasizes interpretability. By projecting the common shift basis back into sensor space, the authors visualize which electrodes contribute most to the shared non‑stationarity. The dominant contributions arise from frontal and central sites, regions known to be sensitive to fatigue, attention, and arousal fluctuations. This neurophysiological insight suggests that the method could be used not only for performance enhancement but also for monitoring user state in adaptive BCI systems.

Limitations are acknowledged. The current formulation assumes linear shift structures; nonlinear dynamics (e.g., sudden drifts or task‑related changes) may not be captured. Real‑time deployment would require efficient online estimation of the subject‑specific shift basis, and the approach has yet to be tested on other modalities such as MEG or fNIRS. Future work could explore kernelized or deep‑learning extensions to model more complex shifts and investigate how the common shift subspace evolves across days or sessions.

In summary, the study introduces a novel paradigm: learning a transferable subspace that encodes common session‑to‑session changes across users, and using it to align a new user’s calibration data with his/her future testing distribution. By decoupling the correction of non‑stationarity from the learning of discriminative features, the method achieves superior classification performance, offers meaningful neurophysiological interpretations, and opens new avenues for robust, user‑friendly BCI deployment.