Auxiliary classifier adversarial networks with maximum subdomain discrepancy for EEG-based emotion recognition.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Domain adaptation (DA) is crucial for unsupervised emotion recognition using electroencephalogram (EEG) data, particularly across different sessions and subjects.
• Existing DA models struggle with cross-domain shifts from individual and session variability, limiting generalization.
• Discrepancies among task-specific subdomains are often overlooked, further impacting model performance.

Purpose of the Study:

• To propose Auxiliary Classifier Adversarial Networks (ACAN) for enhancing unsupervised emotion recognition from EEG.
• To address cross-domain shifts by aligning global and subdomains.
• To maximize subdomain discrepancies for improved model effectiveness.

Main Methods:

• Implemented a domain alignment module in the feature space to minimize inter-domain and inter-subdomain discrepancies.
• Introduced an auxiliary adversarial classifier to generate distinguishable subdomain features via adversarial learning.
• Employed adversarial learning between the feature extractor and the auxiliary classifier.

Main Results:

• ACAN demonstrated effectiveness and superiority in cross-session and cross-subject experiments on SEED, SEED-IV, and DEAP databases.
• The proposed method outperformed state-of-the-art DA techniques in addressing domain shifts.
• Validated the model's ability to enhance emotion recognition accuracy in complex scenarios.

Conclusions:

• ACAN effectively tackles cross-domain shifts in EEG-based emotion recognition.
• The method significantly improves model generalization by aligning domains and subdomains.
• This work advances the development of robust brain-computer interfaces for emotion recognition.