Reducing Calibration Efforts in RSVP Tasks With Multi-Source Adversarial Domain Adaptation.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Brain-Computer Interfaces (BCI) using Rapid Serial Visual Presentation (RSVP) offer efficient information detection.
• Current RSVP-BCI systems require lengthy calibration for new users, limiting practical application.
• Reducing calibration effort is crucial for widespread BCI adoption.

Purpose of the Study:

• To develop a novel framework for reducing calibration data requirements in RSVP-BCI.
• To improve the efficiency and accessibility of BCI systems for new users.
• To leverage existing subject data to train models for new users with minimal input.

Main Methods:

• Proposed a multi-source conditional adversarial domain adaptation with correlation metric learning (mCADA-C) framework.
• Utilized adversarial training for a CNN-based feature extractor to learn common features across domains.
• Incorporated correlation metric learning (CML) loss to enhance feature discrimination.
• Implemented a multi-source framework with a source selection strategy for integrating adaptation results.

Main Results:

• Achieved 87.72% cross-subject balanced-accuracy with only one block of calibration data.
• Demonstrated significant reduction in data requirements for training BCI models.
• Validated the framework on a newly constructed RSVP-based dataset with 11 subjects.

Conclusions:

• The mCADA-C framework effectively reduces calibration efforts in RSVP-BCI systems.
• The proposed method enables higher BCI performance with substantially less user-specific data.
• This advancement promotes more practical and accessible Brain-Computer Interface applications.