Enhancing EEG Classification for Motor Imagery Control of a VR Game based on Deep Learning Techniques on Small Datasets.

Motor imagery-based Brain-Computer Interfaces (BCIs) suffer from limited accuracy when the EEG dataset is recorded from naive BCI users due to noisy components. Neural networks capture more robust representations of EEG features, but require large amount of data which is challenging to collect, due to long motor imagery training sessions. On the other hand, linear- and Riemann-based machine learning algorithms achieve above chance-level accuracy on small scale datasets, but, performance degrades on noisy datasets. To address this issue, we implemented a Wasserstein Generative Adversarial Network (WGAN) for data augmentation to prevent overfitting for the deep classifier, while reaching training convergence faster than existing models. For classification, we developed a Convolutional Neural Network (CNN) to eliminate noisy components caused by BCI illiteracy and extract robust temporal representations of EEG features. To evaluate our system, we designed a VR maze game utilizing the proposed BCI system to translate the EEG signal into movement for a playable character. We achieve increased accuracy, compared to conventional machine learning models, with minimal overfitting, on our own dataset, recorded from 16 naive BCI users.