Cross-Stimulus Transfer Learning: Enhancing Emotion Recognition from Visual-Auditory to Olfactory Perception

Olfactory stimuli, due to their unique neural connections with the amygdala and hippocampus, demonstrate significant potential for emotion research. However, compared to visual-auditory stimuli, they remain relatively underexplored. To address this gap, we designed a cross-stimulus transfer learning task involving two datasets. One dataset contains emotion EEG data induced by olfactory stimuli, while the other, the SEED dataset, contains EEG signals recorded during video-induced emotional experiences. Both datasets focus on three emotional categories: positive, neutral, and negative. To minimize differences between stimulus types and handle individual variability, we employed domain adaptation (DA) and domain generalization (DG) techniques. A Transformer-based Domain-Adversarial Neural Network (DANN) achieved average accuracies of 50.7% and 44.0% in cross-subject and cross-stimulus emotion recognition, respectively, outperforming conventional methods. The study validated the feasibility of cross-stimulus transfer in emotion recognition and provided the experimental support for further exploration of stimulus-independent emotion patterns.