Transfer Learning Enhanced Common Spatial Pattern Filtering for Brain Computer Interfaces (BCIs): Overview and a New Approach

He He,Dongrui Wu
DOI: https://doi.org/10.48550/arXiv.1808.05853
2018-08-09
Abstract:The electroencephalogram (EEG) is the most widely used input for brain computer interfaces (BCIs), and common spatial pattern (CSP) is frequently used to spatially filter it to increase its signal-to-noise ratio. However, CSP is a supervised filter, which needs some subject-specific calibration data to design. This is time-consuming and not user-friendly. A promising approach for shortening or even completely eliminating this calibration session is transfer learning, which leverages relevant data or knowledge from other subjects or tasks. This paper reviews three existing approaches for incorporating transfer learning into CSP, and also proposes a new transfer learning enhanced CSP approach. Experiments on motor imagery classification demonstrate their effectiveness. Particularly, our proposed approach achieves the best performance when the number of target domain calibration samples is small.
Human-Computer Interaction,Computer Vision and Pattern Recognition,Machine Learning
What problem does this paper attempt to address?
The problem this paper attempts to address is: how to reduce or completely eliminate the subject-specific calibration data required for the design of Common Spatial Pattern (CSP) filters in Brain-Computer Interfaces (BCIs) through Transfer Learning (TL). Specifically, CSP is a commonly used spatial filtering method to improve the signal-to-noise ratio of electroencephalogram (EEG) signals, but it is supervised and requires some subject-specific calibration data to design the filters. This process is both time-consuming and user-unfriendly. Transfer Learning can shorten or even completely eliminate this calibration process by leveraging relevant data or knowledge from other subjects or tasks. The main contributions of the paper include: 1. Categorizing existing transfer learning-enhanced CSP methods into two types and providing a comprehensive review. 2. Proposing a new instance-based transfer learning-enhanced CSP method and validating its performance on an EEG-based motor imagery classification task, particularly demonstrating optimal performance when the number of calibration samples in the target domain is small.