Data Augmentation for P300-based Brain-Computer Interfaces Using Generative Adversarial Networks

TL;DR: In this paper, the authors explored the prospect of using Generative Adversarial Networks (GANs) to perform data augmentation for generating artificial training data that are used in the classification of P300 event-related potentials in electroencephalogram (EEG) signals.

Abstract: A notable problem in Brain-Computer Interface (BCI) is the burden of collecting an adequate amount of training data to be used in estimating a robust classifier model. This study explores the prospect of using Generative Adversarial Networks (GANs), a novel family of generative models, to perform data augmentation for generating artificial training data that are used in the classification of P300 event-related potentials in electroencephalogram (EEG) signals. In this regard, we consider two popular GANs, namely, Deep Convolutional Generative Adversarial Networks (DCGAN) and Wasserstein GAN with Gradient Penalty (WGAN-GP), and explore their efficacy in generating artificial EEG data in both subject-specific and subject-independent contexts. The results show that data augmentation using both DCGAN and WGAN-GP could lead to improved classification performance. However, the operating conditions that an improvement is observed differ between the subject-specific and subject-independent classification schemes. In particular, we observe that a transition from a relatively small to large training sample size per subject would generally lead to a better classification performance in training subject-specific classifiers; however, when a limited number of subjects is available, this transition could potentially result in an opposite effect in case of subject-independent classification.