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

Yan Li,Xichen Zhou,Qingguo Wei,Yuebin Chen

doi:10.1109/access.2020.3048683

Yan Li, Xichen Zhou + Show 2 more

Open Access

https://doi.org/10.1109/access.2020.3048683

Copy DOI

Abstract

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

Highlights

A brain-computer interface (BCI) system provides a new non-muscular channel for sending messages from the brain to the external world by analyzing electrical brain activity or other electrophysiological measures of brain functions [1]
Thereby, the data distributions of the raw data, the Riemannian space data alignment (RA) data and the Riemannian and Euclidean space data alignment (REA) data are compared in the Riemannian space, whereas those of the raw data, the Euclidean space data alignment (EA) data and the REA data are compared in Euclidean space
Comparing the two data sets, each algorithm in data set 2 selected more source subjects than its counterpart in data set 1. These results suggest that data alignment does increase the similarity between a source subject and the target subject, and a data set with more subjects is beneficial for transfer learning

Summary

Introduction

A brain-computer interface (BCI) system provides a new non-muscular channel for sending messages from the brain to the external world by analyzing electrical brain activity or other electrophysiological measures of brain functions [1]. As a typical kind of spontaneous systems, motor imagery (MI) based BCIs do not need additional stimulation devices. Instead, their input signals are generated by users via imagining their own limb movements or observing the movements of others [3,4]. In MI based BCIs, the secondorder statistics of EEG signals contain the separable information of brain states [8]. Spatial covariance matrix (SCM) is the most commonly used second-order statistic of EEG signals [9]. A optimization algorithm based on the covariance matrix for either feature extraction or classification is in line with the rationale of MI based BCIs, which are topographyorientated technology.

Methods

Results

Conclusion