bhargavaganti/gan_for_synthetic_eeg_data

Using GAN to create synthetic and partially synthetic EEG data to augment training sets for motor imagery interaction tasks

GAN for Synthetic EEG Data

In this project, we explore the application of a generative adversarial network (GAN) to the production of synthetic EEG data. We examined the benefit of using this synthetic data for dataset augmentation, with the goal of increasing classification accuracy of a SVM in a motor imagery interaction task.

We implement and train a GAN, and run three experiments to evaluate the potential increase in classification accuracy of a logistic regression classifier when trained on an augmented training set rather than a training set consisting of only real EEG data.

Background

In the growing field of research for brain computer interfaces (BCI), EEG data is often measured and classified with various machine learning techniques in order to creative adaptive interfaces based around brain signals. One common problem, which comes from the time consuming nature of collecting EEG data, is a shortage of data available to train models. We explore whether dataset augmentation with GAN could help mitigate this challenge in applying classification algorithms to EEG datasets in the field of BCI.

For more specifics on our model, methods, and results, please refer to our report which can be found in 'report.pdf'.