Domain Adaptation for EEG Emotion Recognition Based on Latent Representation Similarity

摘要

Emotion recognition has many potential applications in the real world. Among the many emotion recognition methods, electroencephalogram (EEG) shows advantage in reliability and accuracy. However, the individual differences of EEG limit the generalization of emotion classifiers across subjects. Moreover, due to the nonstationary characteristic of EEG, the signals of one subject change over time, which is a challenge to acquire models that could work across sessions. In this article, we propose a novel domain adaptation method to generalize the emotion recognition models across subjects and sessions. We use neural networks to implement the emotion recognition models, which are optimized by minimizing the classification error on the source while making the source and the target similar in their latent representations. Considering the functional differences of the network layers, we use adversarial training to adapt the marginal distributions in the early layers and perform association reinforcement to adapt the conditional distributions in the last layers. In this way, we approximately adapt the joint distributions by simultaneously adapting marginal distributions and conditional distributions. The method is compared with multiple representatives and recent domain adaptation algorithms on benchmark SEED and DEAP for recognizing three and four affective states, respectively. The experimental results show that the proposed method reaches and outperforms the state of the arts.