Proceedings of the Annual Meeting of the Cognitive Science Society

Emotion recognition from neural signals constitutes a cornerstone of cognitive science, providing objective access to affective dynamics. However, reliable decoding across datasets is hindered by domain shifts in recording protocols and neurophysiological variability. We propose EmoVAE, a domain adaptation framework that bridges emotional neural dynamics via variational latent distribution alignment. EmoVAE uses a multiscale spatiotemporal aggregation encoder to capture spatiotemporal patterns of affect-related activity, and an EEG-Latent module to map these representations into a continuous probabilistic space. Within this latent space, we align class-conditional distributions using conditional maximum mean discrepancy, yielding stable, geometry-aware emotion clusters across domains. Experiments on SEED and SEED-VII show state-of-the-art performance for both cross-subject and cross-dataset protocols, demonstrating robust generalization under domain shift. These findings suggest that continuous latent alignment is an effective strategy for building reliable EEG-based emotion recognition systems for cognitive and clinical applications.