Online Sequential EEG Emotion Recognition with Prototypical Alignment Based Transfer Model

In recent years, emotion recognition using deep learning has emerged as a popular research topic. However, these models typically require large datasets and struggle to adapt to new data after initial training, limiting further optimization. Additionally, the lack of subject independence in training data often artificially inflates model accuracy. This paper introduces an innovative online sequential electroencephalogram (EEG) emotion recognition method that utilizes a cross-subject transfer learning model within an online learning environment. By selectively pruning and reinitializing model parameters, this approach rapidly adapts to new subjects. Moreover, an enhanced Domain Adversarial Neural Network (DANN) strategy aligns prototype features across emotional categories within the transfer learning framework, thereby improving model accuracy and simplifying the network architecture. Experiments conducted on the SEED and SEED-IV datasets illustrate that the proposed method yields average accuracies of 80.04% and 62.78%, respectively, surpassing existing mainstream online learning methods. The study also explores how the quantity of pre-trained subjects impacts the accuracy in predicting new subjects in an online learning scenario. Results indicate that this method can quickly adapt to new subjects under limited sample conditions while maintaining high accuracy. These findings highlight the practical application potential of this approach across fields such as neuroscience, computer science and human-computer interaction.