PARSE: Pairwise Alignment of Representations in Semi-Supervised EEG Learning for Emotion Recognition

We propose PARSE, a novel semi-supervised architecture for learning strong EEG representations emotion recognition. To reduce the potential distribution mismatch between large amounts of unlabeled data and limited amount labeled data, PARSE uses pairwise representation alignment. First, our model performs augmentation followed by label guessing original augmented data. This is then sharpening guessed labels convex combinations Finally, alignment classification are performed. rigorously test model, we compare to several state-of-the-art approaches which implement adapt learning. perform these experiments on four public EEG-based recognition datasets, SEED, SEED-IV, SEED-V AMIGOS (valence arousal). The show that proposed framework achieves overall best results with varying samples in SEED-IV (valence), while approaching result (reaching second-best) (arousal). analysis shows considerably improves performance reducing especially when only 1 sample per class labeled.