PC-SSL: A Predictive Coding-Based Self-Supervised Learning Framework for EEG Emotion Recognition

We introduce a predictive coding self-supervised learning (PC-SSL) approach to overcome the limitations of traditional supervised methods hindered by the high cost of labeling EEG data and label scarcity. PC-SSL is used for emotion recognition from EEG signals leveraging unlabeled data. The predictive coding framework learns to forecast future EEG representations based solely on past neural activity. Our model employs a convolutional encoder augmented with band-wise and channel-wise attention mechanisms to extract local spatio-spectral features from differential entropy (DE) representations of EEG signals. Evaluated on the SEED-IV and SEED-V datasets, PC-SSL attains 84.48% and 92.39% accuracy, respectively — improving upon the most recent baselines by approximately 18% and 16%.