Multisource Transfer Learning for Cross-Subject EEG Emotion Recognition.

Area of Science:

• Neuroscience
• Machine Learning
• Affective Computing

Background:

• Electroencephalogram (EEG) is crucial for emotion recognition due to its temporal resolution.
• Individual EEG differences necessitate personalized models, increasing data collection demands.
• Fast deployment of emotion recognition models requires reducing reliance on labeled data.

Purpose of the Study:

• To develop a multisource transfer learning method for efficient EEG-based emotion recognition.
• To enable rapid model adaptation for new users with minimal labeled data.
• To address the challenge of inter-person variability in EEG signals.

Main Methods:

• Proposed a multisource transfer learning framework with existing users as sources and new users as targets.
• Implemented source selection to identify relevant existing models.
• Utilized style transfer mapping to minimize EEG differences between target and source domains.
• Employed few-shot learning in calibration sessions for source selection and style transfer.

Main Results:

• Achieved a 12.72% improvement in three-category classification accuracy on the SEED benchmark compared to non-transfer methods.
• Demonstrated effective reduction of EEG inter-person differences through style transfer.
• Validated the method's efficacy in fast-deployment scenarios.

Conclusions:

• The proposed multisource transfer learning method significantly enhances EEG-based emotion recognition performance for new users.
• This approach effectively reduces the need for large amounts of labeled data, facilitating rapid deployment.
• The method holds practical significance for real-world applications requiring quick adaptation of emotion recognition models.