A robust semi-supervised deep learning approach for emotion recognition using EEG signals

Many deep learning models are recently proposed for Electroencephalography (EEG) classification tasks. However, they are full-supervised and require large amounts of labeled data. Labeling EEG signals is a time-consuming and expensive process needing many trials and careful analysis by the experts. Recently, many modern semi-supervised methods are proposed that require less supervised information to achieve competitive performance with that of supervised ones, but they are mainly developed in the computer vision domain and adapting these methods for EEG applications is an open issue. This paper presents a robust semi-supervised deep Learning method. To this end, we design appropriate augmentations for EEG signals leading to promising results in a low-supervised setting. Especially, compared to naïve Gaussian noise used in previous work, the proposed strong augmentation boosts the performance of our method by a large margin. We also enhance our method by utilizing distribution alignment and relative confidence threshold techniques. We carry out several experiments on the Database for Emotion Analysis using Physiological dataset in both valence/arousal emotion recognition tasks. The results confirm that the proposed method leverage the unlabeled information effectively and significantly outperforms the peer methods.