Answers are grounded in this paper and cite their sources.
Objective. This study compares Riemannian Procrustes analysis (RPA) and tangent space alignment (TSA) for transfer learning of motor-imagery EEG data. The study also examines how performance changes when the EEG data is represented using analytic Hermitian positive definite covariance matrices instead of standard symmetric positive definite covariance matrices. Approach. The methods' geometric constraints were compared and explained. Additionally, the methods were evaluated in two scenarios on real EEG data, one with substantial target-domain knowledge and one calibration-like with a limited amount of training data and thus limited target-domain knowledge. Six public EEG datasets with multiple subjects performing two or multiple motor imagery tasks were used. We report paired comparisons of classification performance for the different transfer learning methods and covariance matrix representations across the datasets in order to draw general conclusions about performance. Main results. In the scenario with substantial target-domain knowledge, the TSA method outperformed the RPA method, and representing the EEG data as analytical covariance matrices outperformed standard covariance matrices. In the scenario with limited target-domain knowledge, performance depended on the amount of available target data, and source-data selection emerged as an important component of successful transfer learning. Significance. This study gives a deeper understanding of TSA and RPA transfer learning methods commonly used in brain-computer interfaces, which may inform future research and implementation. The analysis also demonstrates the importance of source data selection in brain-computer interface transfer learning.
No takes yet. Share an insight, caveat, or question.
Nilsson et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: