Prompt Learning-based Pretrained Visual Transformer for Emotion Recognition Using Electroencephalogram

Emotion recognition using electroencephalogram (EEG) is a vital problem in affective computing and human-computer interaction applications. By recognizing and classifying the provided EEG signals into different categories, a subject's emotional state can be measured and modeled during human-computer emotional interactions. Notwithstanding the fact that earlier deep learning techniques have demonstrated remarkably high accuracy in addressing this task, the inter-subject variability in EEG still significantly affects the accuracy of recognition, which needs to be improved. In this paper, we propose a prompt learning-based pretrained visual Transformer, abbreviated as PLPViT, for EEG-based cross-subject emotion recognition tasks. The optimization and inference are conducted on the SEED dataset, a public benchmark dataset that contains three different emotion types: positive, neutral, and negative. The proposed method achieved a classification accuracy of 83.48%, a competitive result compared to other methods under cross-subject experiment protocols, demonstrating the feasibility of prompt techniques and pretrained large models for EEG-based cross-subject emotion recognition tasks.