A Manifold Learning Approach to Chart Human Brain Dynamics Using Resting EEG Signals

Abstract

In this study, we propose an approach to identify individuality that appears in human brain dynamics and visualize inter-individual variations in a low-dimensional space. For this purpose, we first introduce an appropriate similarity measure between multichannel electroencephalography (EEG) signals based on information geometry. Then, we use t-distributed stochastic neighbor embedding, which is a state-of-the-art algorithm for manifold learning, and combine it with the information distance to map points in the high-dimensional EEG signal space into two-dimensional space. We show that a fine low-dimensional visualization enables us to identify each subject as an isolated island of points and preserve inter-individual variations. We also provide an appropriate approach to tune the cost function parameter.