Learning Subject-Generalized Topographical EEG Embeddings Using Deep Variational Autoencoders and Domain-Adversarial...

Area of Science:

• Neuroscience
• Machine Learning
• Affective Computing

Background:

• Emotion detection from electroencephalography (EEG) faces challenges due to limited labeled data and significant inter-subject variability.
• Existing models struggle to generalize across diverse user populations, hindering practical applications.

Purpose of the Study:

• To develop a context-generalized deep neural network model for robust emotion detection from EEG data.
• To simultaneously address data scarcity and subject variability in EEG-based emotion recognition.

Main Methods:

• Utilized variational autoencoders (VAEs) for training on both labeled and unlabeled data, maximizing data utilization.
• Implemented variational regularization to promote Gaussian-distributed, subject-independent feature embeddings.
• Applied subject-adversarial regularization to bi-lateral features for enhanced subject-independence in emotion classification.

Main Results:

• The proposed model demonstrated superior subject-independent generalization performance compared to state-of-the-art methods on SEED and DEAP datasets.
• Qualitative analysis indicated that the subject-invariant bi-lateral variational domain adversarial neural network (BiVDANN) architecture discovers normally distributed features, improving cross-subject performance.

Conclusions:

• The BiVDANN architecture effectively tackles data constraints and subject variability in EEG emotion detection.
• The model offers improved generalization capabilities for real-world applications of affective computing using EEG.