LAUSR

In the rapid serial visual presentation (RSVP) classification task, the data from the target and non-target classes are incredibly imbalanced. These class imbalance problems (CIPs) can hinder the classifier from achieving better performance, especially in deep learning. This paper proposed a novel data augmentation method called balanced Wasserstein generative adversarial network with gradient penalty (BWGAN-GP) to generate RSVP minority class data. The model learned useful features from majority classes and used them to generate minority-class artificial EEG data. It combines generative adversarial network (GAN) with autoencoder initialization strategy enables this method to learn an accurate class-conditioning in the latent space to drive the generation process towards the minority class. We used RSVP datasets from nine subjects to evaluate the classification performance of our proposed generated model and compare them with those of other methods. The average AUC obtained with BWGAN-GP on EEGNet was 94.43%, an increase of 3.7% over the original data. We also used different amounts of original data to investigate the effect of the generated EEG data on the calibration phase. Only 60% of original data were needed to achieve acceptable classification performance. These results show that the BWGAN-GP could effectively alleviate CIPs in the RSVP task and obtain the best performance when the two classes of data are balanced. The findings suggest that data augmentation techniques could generate artificial EEG to reduce calibration time in other brain-computer interfaces (BCI) paradigms similar to RSVP.