GAN-Based Data Augmentation For Improving The Classification Of EEG Signals

Abstract

Emotion recognition is a field of psychology that involves the process of identifying emotions and treating mental conditions like autism. The advancements in the field of machine learning and deep learning have paved the way for scientists to develop models for evaluating emotions by analyzing facial expressions, speech and text. However, the task of evaluating emotions could be best done by processing the bio-signals and neural imaging of the brain. In that sense, bio-signals such as Electroencephalogram (EEG) are less expensive to use and non-invasive, giving them an edge over traditional methods like Magnetic Resonant Imaging (MRI). However, not many datasets are publicly available due to privacy issues and their availability is highly limited by the classification task. These constraints, along with the problem of data scarcity, motivates this work as an attempt to enhance the accuracy scores by generating synthetic features that are close to actual data distribution. In this research, we propose a Wasserstein Generative Adversarial Network with gradient penalty (WGAN-GP) based model that can help tackle this problem. The dataset that is investigated is DEAP, one of the benchmark datasets for evaluating emotion recognition algorithms. In the method proposed, nine descriptive features are extracted from the original data and baseline models are evaluated. Subsequently, a WGAN-GP is trained on these extracted features and it is used to generate a new set of synthetic data features. The synthetic features are then analysed for quality and appended to the original data to expand this dataset. Experiments with different augmentation factors (x2, x3, x4) are investigated to evaluate the impact of the data augmentation procedure. The experimental results demonstrate that the proposed method gives a considerable enhancement of the classification task’s performance.