A Transfer Learning Framework for RSVP-based Brain Computer Interface^.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Rapid Serial Visual Presentation (RSVP)-based Brain-Computer Interfaces (BCI) offer efficient information detection via brain responses to visual stimuli.
• Current BCI systems require lengthy calibration for new users, hindering practical application.
• Reducing calibration time is crucial for widespread BCI adoption.

Purpose of the Study:

• To develop a novel multi-source transfer learning framework to minimize data requirements for new users in RSVP-BCI systems.
• To address the challenge of time-consuming calibration in electroencephalogram (EEG)-based BCIs.
• To enhance the efficiency and accessibility of BCI technology.

Main Methods:

• Collected an RSVP-based EEG dataset from 11 subjects performing an image retrieval task.
• Proposed a transfer learning framework incorporating a source-selection strategy to prevent negative transfer.
• Developed a domain adversarial training network using Convolutional Neural Networks (CNNs) to extract common EEG features across subjects, with a classifier for semantic information and techniques for stable adversarial training.

Main Results:

• The proposed transfer learning method significantly reduced the data needed for new subject calibration.
• The domain adversarial network effectively learned common EEG features while distinguishing between subjects.
• Experimental results showed superior performance compared to existing state-of-the-art and baseline approaches.

Conclusions:

• The developed multi-source transfer learning framework effectively reduces calibration effort in RSVP-BCI systems.
• This approach enhances the practicality and usability of EEG-based BCI technology for new users.
• The study demonstrates a promising direction for efficient and personalized BCI calibration.