Hybrid self-supervised EEG emotion representation: masked reconstruction joint with mutual information bounds

Area of Science:

• Neuroscience
• Machine Learning
• Affective Computing

Background:

• Electroencephalography (EEG) is crucial for non-invasive emotion recognition in affective computing.
• Deep learning and self-supervised learning (SSL) are vital for EEG representation learning.
• Existing SSL methods struggle to balance structural fidelity and semantic discriminability in EEG signals.

Purpose of the Study:

• To propose HybridMI-EEG, a novel hybrid self-supervised framework for EEG emotion recognition.
• To address the limitations of existing methods in balancing structural and semantic feature extraction.
• To enhance the performance and stability of EEG-based emotion classification.

Main Methods:

• Developed a hybrid self-supervised framework (HybridMI-EEG) utilizing a Transformer encoder-decoder architecture.
• Integrated masked reconstruction for capturing structural information and MI bound optimization for semantic feature extraction.
• Employed a dual-objective loss function and learnable augmenters for synergistic optimization of local structure and global semantics.

Main Results:

• HybridMI-EEG demonstrated superior performance compared to state-of-the-art baselines on DEAP and DREAMER datasets.
• The framework achieved higher accuracy and improved stability in EEG emotion recognition.
• The proposed method effectively balances structural fidelity and semantic discriminability in EEG representations.

Conclusions:

• HybridMI-EEG offers an effective solution for enhancing EEG representation learning and emotion classification.
• The framework provides a promising approach for advancing affective computing and brain informatics.
• The synergistic optimization of local structure and global semantics is key to improved EEG emotion recognition.