Emotion Recognition Based on Electroencephalogram Using Semi-supervised Generative Adversarial Network.

Abstract

This paper proposes the use of Semi-supervised Generative Adversarial Network (SGAN) to take advantage of the large amount of unlabeled electroencephalogram (EEG) spectrogram data in improving the classifier's accuracy in emotion recognition. The use of SGAN led the discriminator network to not just learn in a supervised fashion from the small amount of labeled data to distinguish among the different target classes, but also make use of the true unlabeled data to distinguish them from the synthetic ones generated by the generator network. This additional ability to distinguish true and fake samples forces the network to focus only on features that are present on a true sample to distinguish the classes, thereby improving generalization and overall accuracy. An ablation study is devised, where the SGAN classifier is compared to a mere discriminator network without the GAN architecture. The 80% : 20% validation method was employed to classify the EEG spectrogram of 50 participants gathered by Kaohsiung Medical University into two emotion labels in the valence dimension: positive and negative. The proposed method achieved an accuracy of 84.83% given only 50% labeled data, which is not just better than the baseline discriminator network which achieved 83.5% accuracy, but is also better than many previous studies at accuracies around 78%. This demonstrates the ability of SGAN in improving discriminator network's accuracy by training it to also distinguish between the unlabeled true sample and synthetic data.Clinical Relevance- The use of EEG in emotion recognition has seen growing interest due to its ease of access. However, the large amount of labeled data required to train an accurate model has been the limiting factor as databases in the area of emotion recognition with EEG is still relatively small. This paper proposes the use of SGAN to allow using large amount of unlabeled EEG data to improve the recognition rate.