Enhancing the Classification of EEG Signals using Wasserstein Generative Adversarial Networks Vlad Mihai Petrut ¸iu Technical University of Cluj-Napoca Cluj-Napoca, Romania Email: petrutiuvlad@yahoo.ro Liana Daniela Palcu Technical University of Cluj-Napoca Cluj-Napoca, Romania Email: ldpalcu@gmail.com Camelia Lemnaru Technical University of Cluj-Napoca Cluj-Napoca, Romania Email: Camelia.Lemnaru@cs.utcluj.ro Mihaela Dˆ ıns ¸oreanu Technical University of Cluj-Napoca Cluj-Napoca, Romania Email: Mihaela.Dinsoreanu@cs.utcluj.ro Rodica Potolea Technical University of Cluj-Napoca Cluj-Napoca, Romania Email: Rodica.Potolea@cs.utcluj.ro Raul Mures ¸an Transylvanian Institute of Neuroscience Cluj-Napoca, Romania Email: muresan@tins.ro Vlad Vasile Moca Transylvanian Institute of Neuroscience Cluj-Napoca, Romania Email: moca@tins.ro Abstract—Collecting EEG signal data during a human vi- sual recognition task is a costly and time-consuming process. However, training good classification models usually requires a large amount of quality data. We propose a data augmentation method based on Generative Adversarial Networks (GANs) to generate artificial EEG signals from existing data, in order to improve classification performance on data collected during a visual recognition task. We evaluate the quality of the artificially generated signal in terms of the accuracy of a Convolutional Neural Network-based classifier that uses both real and aug- mented data to classify the outcome of the cognitive task. The preliminary results suggest that the introduction of artificially generated signals have a positive effect on the performance of the classifier. Moreover, we provide a method to quantify the level of information which indicates that the generated signals indeed follow the properties of the real ones. I. I NTRODUCTION AND PROBLEM STATEMENT It is widely known that Machine Learning models require a large quantity of high-quality data in order to produce state of the art results. Neuroscience is the domain which constitutes both a source of inspiration, but also a very challenging, yet attractive application domain for Deep Learning methods. One downside when applying Deep Learning strategies on Neuroscience data is that the data collection process is very costly and time consuming. Also, depending on the technique used during collection, the signals might be affected by external factors and thus they might contain unwanted noise. Additionally, the ability of the algorithm to generalize on this domain is generally limited because the variability of neural activity and interaction is very high among individuals. For addressing all these problems, this paper proposes the use of a data augmentation technique. Data augmentation has proved to be extremely useful in adding variety to the original data and multiplying the number of examples which were originally captured. The Generative Adversarial Networks (GANs) [1] were first developed in the Computer Vision domain for producing new examples which respected an initial distribution of the features, but were extremely different from the original ones. Following its initial success, the method started to be expanded and applied in other domains too. One of the goals when analysing brain signals is to gain insight into how neurons communicate while performing a cognitive task, such as object recognition. One step in this process is to build models which try to capture the real interaction between neurons or neural areas. However, it is difficult to evaluate whether the models encode indeed the neural interaction. An indirect solution to this is to further use that information in a classification task. Additionally, the hypothesis is that - by appropriately analysing the knowledge encoded in those models - one might achieve a better under- standing of how the brain processes visual information. To that end, this paper explores the possibility of using Wasserstein GANs (WGANs) [2] for generating synthetic signals for enhancing the classification results on a visual recogition task, using data from ElectroEncephaloGram (EEG) recordings. II. STATE OF THE ART Initially explored in the Computer Vision domain, Gener- ative Adversarial Networks [1] have quickly replaced Varia- tional Auto-Encoders (VAEs) [3] as data augmentation strate- gies, as they were found to provide better variety in the training data, thus boosting the performance of learning algorithms. Initially, VAEs managed to capture relevant features for low- sized images, but did not manage to focus on details, this being 978-1-7281-9080-8/20/$31.00 c 2020 IEEE