Valence-Arousal Disentangled Representation Learning for Emotion Recognition in SSVEP-Based BCIs

Area of Science:

• Neuroscience
• Human-Computer Interaction
• Machine Learning

Background:

• Steady state visually evoked potential (SSVEP)-based brain-computer interfaces (BCIs) are vital for rehabilitation and disability assistance.
• Real-time emotion recognition can significantly improve human-machine interaction in SSVEP-BCIs.
• Current methods face challenges with generalization and accuracy in emotion detection due to unintended latent representation learning.

Purpose of the Study:

• To introduce a novel Valence-Arousal Disentangled Representation Learning (VADL) method for SSVEP-BCIs.
• To enhance the performance and generalization of emotion recognition within SSVEP-BCIs.
• To improve the accuracy of detecting emotional states by disentangling valence and arousal information.

Main Methods:

• Developed a Valence-Arousal Disentangled Representation Learning (VADL) method inspired by the two-dimensional emotional model.
• Utilized a structured state space duality model for comprehensive global emotional feature extraction.
• Implemented a Multisubject Gradient Blending training strategy to tailor learning paces for reconstruction and discrimination tasks.

Main Results:

• VADL effectively disentangles latent variables for valence and arousal, improving accuracy.
• The structured state space duality model successfully extracted global emotional features.
• Experimental results demonstrate that VADL outperforms existing state-of-the-art benchmark algorithms.
• A comprehensive database with 23 subjects was created to validate the method.

Conclusions:

• The VADL method significantly enhances emotion recognition accuracy and generalization in SSVEP-BCIs.
• The proposed Multisubject Gradient Blending strategy aids in adaptive learning for individual subjects.
• This research offers a promising advancement for more intuitive and effective human-machine interaction in BCI applications.