Self supervised learning based emotion recognition using physiological signals

Introduction: Emotion recognition is important for human–machine interaction. Physiological-signal-based methods, especially EEG, can objectively reflect emotional states, but labeled EEG datasets are typically small. Methods: Large amounts of unlabeled EEG are easier to obtain, so this work adopts self-supervised learning (SSL) for emotion recognition from EEG. Three pretext tasks are used to define pseudo-labels and extract features from the data’s inherent structure: Contrastive Predictive Coding (CPC), Relative Position (RP), and Temporal Shuffling (TS). Results and discussion: Experiments show that SSL can learn effective feature representations for downstream emotion recognition without manual labels, demonstrating its potential for this problem.