Cross-subject Emotion Recognition And Vigilance Estimation With Adversarial Domain Generalization

Due to the subject variability between the physiological signals of different subjects,the brain-computer interface research has faced great obstacle when applying to large-scale real-world scenarios.On one hand,the traditional machine learning methods suffer from performance degradation since the different distribution between the training and the testing subjects.On the other hand,existing methods focus on domain adaptation,providing personalized solution for new users.However,the training and calibration of the personalized model still needs timeconsuming and expensive data recollection,which makes the approach inefficient and impractical.In fact,the cross-subject model is supposed to obtain enough generalization ability for new users in the depersonalized conditions,where no information from the new users is available during training.In this paper,we manage to reduce the influence of subject variability by introducing domain generalization to cross-subject braincomputer interface research.Since the adversarial learning structure of Domain Adversarial Neural Network(DANN)has achieved great improvement on cross-subject models,we generalized the DANN to domain generalization problem,and then proposed a novel model,the Domain Residual Network,by further modification.Two different datasets have been utilized to evaluate the proposed methods,including 15-subject three emotion classification problem on SEED and 23-subject vigilance estimation regression problem on SEED-VIG.Experimental results have shown that when applying to one unknown subject,the proposed method based on domain generalization is able to achieve similar performance with the domain adaptation models which are trained with unlabeled data from the target subject.Additionally,other results also indicated that the proposed models are effective when dealing with multiple previously unseen subjects.The conclusions of this research have provided a feasible approach and potential directions for depersonalized cross-subject brain-computer interface systems in real-world applications.