Transfer Learning Based on Hybrid Riemannian and Euclidean Space Data Alignment and Subject Selection in Brain-Computer Interfaces

Transfer learning is a promising approach for reducing training time in brain-computer interface (BCI). However, how to effectively transfer data from previous users new user poses huge challenge. This paper presents novel that combines alignment and source subject selection motor imagery (MI) based BCIs. The former achieved by reference matrix the regularization of two matrices estimated Riemannian Euclidean space respectively, whereas latter implemented modified sequential forward floating-point search algorithm. aligned chosen subjects are used creating classification model on either spatial covariance or common pattern algorithm space. proposed algorithms were evaluated MI BCI sets with different compared existing sole selection. experimental results show hybrid-space methods differences among significantly outperform single-space methods, method can substantially enhance similarity between target subject. combination achieves superior performance one. will greatly facilitate real-world applications