Abstract
Non-invasive brain-computer interfaces (BCIs) provide us with the unique ability to classify the psychological state of a person using only neurophysiological signals, such as those captured with an electroencephalogram (EEG). With this ability, new avenues for innovation in the field of healthcare arise, especially as it is used for robotics. EEGNet is a novel deep learning technique for the classification of EEG data with a limited training set that generalizes well to a variety of BCI paradigms, and the performance thereof can further be improved. We propose the use of Thomson Multitaper Power Spectral Density estimation in the EEG-BCI classification pipeline as well as a novel convolutional neural network (CNN), which extends EEGNet with sparse feature maps produced by efficient regularized separable convolutions. Further, we test the efficacy of interspersed Gaussian noise as a data augmentation technique. To show the improvements found with this new pipeline, we test on a widely used public EEG dataset related to emotion classification, then perform an ablation study to determine the most contributing factors. The accuracy on this public dataset was 77.16%. These results show that our pipeline improved the classification accuracy by 10.86% when compared with the state-of-the-art.