Clustering brain-network time series by Riemanninan geometry

SpeakerDr. Kostas Slavakis, Ph.D.
Organization Assistant Professor, Department of Electrical Engineering, University at Buffalo
LocationEBIII 2213
DateNovember 3, 2017 11:45 AM

Abstract: We claim that the latent Riemannian geometry of brain-network time series is beneficial to clustering states of the brain. To this end, two features-extraction schemes which reveal the underlying Riemannian geometry are demonstrated. The first one uses ARMA modeling to generate features/points on the Grassmann manifold, while the second one computes partial correlations to yield points on the manifold of positive definite matrices. Each state of the brain is viewed as a cluster, and each cluster is modeled as a submanifold of the Riemannian manifold at hand. Clustering amounts to the task of segmenting multiple Riemannian Submanifolds. A clustering algorithm is also introduced with theoretical guarantees of performance in one of its special forms. Extensive numerical tests on synthetic and real brain-network data show that the proposed Riemannian-geometry-cognizant framework outperforms classical and state-of-the-art clustering techniques.

Bio: Kostas Slavakis did his graduate (PhD 2002) and post-graduate studies in electrical and electronic engineering at Tokyo Institute of Technology, Tokyo, Japan. He is currently an Associate Professor at the Electrical Engineering Dept. of University at Buffalo, The State University of New York. He has served as a tenured Assist. Prof. at the University of Peloponnese, Greece, and as a Research Assoc. Prof. at the University of Minnesota, USA.

He was also an Assoc. and Senior Area Editor of the IEEE Transactions on Signal Processing, He currently serves as an Assoc. Editor of the EURASIP Signal Processing Journal and the Journal on Advances in Signal Processing. His current research interests lie in the areas of signal processing, machine learning and data analytics.

  November 2017
Sun Mon Tues Wed Thu Fri Sat