Neuron Perception Inspired EEG Emotion Recognition With Parallel Contrastive Learning

Area of Science:

• Neuroscience
• Machine Learning
• Affective Computing

Background:

• Interindividual variability in electroencephalogram (EEG) signals poses significant challenges for subject-independent emotion recognition.
• Existing cross-subject EEG emotion recognition methods lack sufficient exploration of shared neural underpinnings of affective processing.

Purpose of the Study:

• To propose a novel Parallel Contrastive Multisource Domain Adaptation (PCMDA) model for subject-independent EEG-based emotion recognition.
• To address the limitations in uncovering shared neural mechanisms for emotion processing across subjects.
• To improve the generalization capability of EEG emotion recognition models.

Main Methods:

• Developed a PCMDA model inspired by neural representation mechanisms in the ventral visual cortex.
• Employed a neuron-perception-inspired contrastive learning architecture with a two-stage domain alignment methodology.
• Integrated a parallel contrastive loss (PCL) simulating brain's self-supervised learning and a self-attention mechanism for frequency band emotion weighting.

Main Results:

• Extensive experiments on SEED, DEAP, and FACED datasets demonstrated the effectiveness of the PCMDA model.
• The PCMDA model successfully utilized unique EEG features and frequency band information specific to each subject.
• Achieved improved generalization performance across different subjects compared to existing methods.

Conclusions:

• The proposed PCMDA model offers a significant advancement in subject-independent EEG emotion recognition.
• The model's ability to align multisource domains and leverage subject-specific features enhances cross-subject generalization.
• This approach provides a promising direction for understanding and recognizing emotions from EEG data universally.