Using Deep Learning to Classify Saccade Direction from Brain Activity

Ard Kastrati, Martyna Beata Plomecka, Roger Wattenhofer, Nicolas Langer

Abstract: We present first insights into our project that aims to develop an Electroencephalography (EEG) based Eye-Tracker. Our approach is tested and validated on a large dataset of simultaneously recorded EEG and infrared video-based Eye-Tracking, serving as ground truth. We compared several state-of-the-art neural network architectures for time series classification: InceptionTime, EEGNet, and investigated other architectures such as convolutional neural networks (CNN) with Xception modules and Pyramidal CNN. We prepared and tested these architectures with our rich dataset and obtained a remarkable accuracy of the left/right saccades direction classification (94.8 %) for the InceptionTime network, after hyperparameter tuning.