Cross-subject emotion recognition with loop adaptive adversarial transfer network.

Area of Science:

• Neuroscience
• Computer Science
• Artificial Intelligence

Background:

• Electroencephalogram (EEG) signals hold promise for emotion recognition.
• Subject variability in EEG signals presents a significant challenge for practical applications.
• Existing methods struggle with cross-subject generalization in EEG-based emotion recognition.

Purpose of the Study:

• To propose a novel Loop Adaptive Adversarial Transfer Network (LATN) for enhanced cross-subject emotion recognition.
• To address the challenge of individual differences in EEG signals for emotion recognition.
• To improve the accuracy and robustness of EEG-based emotion recognition systems across different subjects.

Main Methods:

• Developed a Loop Adaptive Adversarial Transfer Network (LATN) incorporating Structure-aware Associative Alignment (SAA) and an Inner and Outer Product Combination Strategy (IOPC).
• Employed a semi-supervised learning approach with judicious pseudo-labeling to mitigate negative transfer issues.
• Validated the LATN model on public (DEAP, SEED) and private (ECPL) datasets.

Main Results:

• Achieved 96.33% three-class accuracy on the ECPL dataset.
• Obtained 89.21% binary classification accuracy for arousal and 76.12% for valence on the DEAP dataset.
• Reached 94.34% three-class accuracy on the SEED dataset, demonstrating state-of-the-art performance.

Conclusions:

• The proposed LATN effectively enhances cross-subject emotion recognition from EEG signals.
• LATN demonstrates superior performance compared to existing algorithms across multiple datasets.
• This approach offers a promising solution for real-world applications of EEG-based emotion recognition.