[Paper Review] Transfer Learning for EEG-Based Brain-Computer Interfaces: A Review of Progress Made Since 2016

Dongrui Wu, Yifan Xu+1 more|arXiv (Cornell University)|Apr 13, 2020

EEG and Brain-Computer Interfaces117 references17 citations

TL;DR

This paper reviews transfer learning (TL) approaches in EEG-based brain-computer interfaces (BCIs) since 2016, focusing on cross-subject, cross-session, cross-device, and cross-task transfer across six paradigms: motor imagery, event-related potentials, steady-state visual evoked potentials, affective BCIs, regression problems, and adversarial attacks. It identifies Riemannian geometry and deep learning as emerging key techniques, highlights underexplored areas like cross-task transfer, and emphasizes integration of TL with other methods like active learning for improved performance and reduced calibration needs.

ABSTRACT

A brain-computer interface (BCI) enables a user to communicate with a computer directly using brain signals. The most common non-invasive BCI modality, electroencephalogram (EEG), is sensitive to noise/artifact and suffers between-subject/within-subject non-stationarity. Therefore, it is difficult to build a generic pattern recognition model in an EEG-based BCI system that is optimal for different subjects, during different sessions, for different devices and tasks. Usually, a calibration session is needed to collect some training data for a new subject, which is time-consuming and user unfriendly. Transfer learning (TL), which utilizes data or knowledge from similar or relevant subjects/sessions/devices/tasks to facilitate learning for a new subject/session/device/task, is frequently used to reduce the amount of calibration effort. This paper reviews journal publications on TL approaches in EEG-based BCIs in the last few years, i.e., since 2016. Six paradigms and applications -- motor imagery, event-related potentials, steady-state visual evoked potentials, affective BCIs, regression problems, and adversarial attacks -- are considered. For each paradigm/application, we group the TL approaches into cross-subject/session, cross-device, and cross-task settings and review them separately. Observations and conclusions are made at the end of the paper, which may point to future research directions.

Motivation & Objective

• To systematically review recent advances in transfer learning (TL) for EEG-based brain-computer interfaces (BCIs) since 2016.
• To analyze TL approaches across six BCI paradigms: motor imagery, event-related potentials, steady-state visual evoked potentials, affective BCIs, regression problems, and adversarial attacks.
• To evaluate the effectiveness of TL in cross-subject, cross-session, cross-device, and cross-task settings to reduce calibration burden in EEG-BCI systems.
• To identify research gaps and emerging trends, particularly in underexplored areas like cross-task transfer and integration with active learning.
• To provide insights into future directions for developing more robust, generalizable, and user-friendly EEG-BCI systems.

Proposed method

• The paper conducts a systematic review of journal publications on TL in EEG-BCIs from 2016 onward, categorizing approaches by transfer setting: cross-subject/session, cross-device, and cross-task.
• For each BCI paradigm, TL methods are grouped and analyzed based on their core techniques, such as domain adaptation, metric learning, and deep neural networks.
• Key techniques reviewed include Riemannian geometry for covariance matrix modeling, deep learning architectures like EEGNet and Deep ConvNet, and signal processing methods such as transfer kernel common spatial patterns (TK-CSP).
• The review evaluates methods for trial alignment (e.g., Riemannian alignment, exemplar alignment), feature selection (e.g., CSDF-ReliefF, CSDF-mRMR), and integration with signal filtering and representation learning.
• The paper also examines TL in adversarial settings, leveraging the transferability of adversarial examples across models and domains.
• Integration of TL with other machine learning techniques, such as active learning, is discussed as a pathway to further reduce calibration data requirements.

Experimental results

Research questions

• RQ1 How have transfer learning approaches evolved in EEG-based BCIs since 2016 across different paradigms and transfer settings?
• RQ2 What are the most effective TL techniques for reducing calibration time in cross-subject and cross-session EEG-BCI applications?
• RQ3 To what extent can TL be applied to emerging BCI applications such as affective BCIs and regression-based user state estimation?
• RQ4 Why is cross-task transfer in EEG-BCIs largely underexplored, and what are the potential benefits of developing such methods?
• RQ5 How can TL be effectively combined with adversarial attack frameworks to improve attack transferability in EEG-BCI systems?

Key findings

• Cross-subject and cross-session transfer remain the most studied settings, with significant progress in domain adaptation and deep learning-based methods.
• Riemannian geometry and deep learning have emerged as leading approaches, particularly for modeling EEG covariance matrices and learning invariant representations.
• Only one study on cross-task transfer was identified since 2016, indicating a major research gap in this area.
• Adversarial attacks in EEG-BCIs benefit from transferability of adversarial examples, and explicitly incorporating TL can enhance attack success rates in black-box scenarios.
• TL can be applied beyond classification, including in signal filtering (e.g., TK-CSP), feature selection (e.g., CSDF-ReliefF), and trial alignment (e.g., RA, EA), improving overall pipeline performance.
• Integration of TL with active learning has been shown to further reduce calibration data needs, suggesting a promising hybrid approach for future BCI systems.