Label-free Domain Adaptation and Multimodal Manifold Learning with Riemannian Geometry. Ronen Talmon (Technion)

Title:
Label-free Domain Adaptation and Multimodal Manifold Learning with Riemannian Geometry
Abstract:
Recently, Riemannian geometry has become a central ingredient in a broad range of data analysis and learning tools. Broadly, it facilitates features of complex high-dimensional data with a known non-Euclidean geometry. In this talk, we will first consider the Riemannian geometry of symmetric positive-definite matrices and propose a method based on Procrustes analysis for label-free domain adaptation. We will present some theoretical guarantees and demonstrate the performance on simulations as well as on real-measured data. While most of the talk will be focused on a particular geometry of symmetric positive-definite matrices, we will present generalizations to other geometries (e.g., to hyperbolic spaces) and features (e.g., kernels and graph-Laplacians). Finally, we will introduce some intriguing spectral properties and demonstrate their application to multimodal manifold learning.

Speaker’s Short bio:
Ronen Talmon is an Associate Professor of electrical and computer engineering at the Technion — Israel Institute of Technology. His research interests are geometric signal and data processing, manifold learning, and applied harmonic analysis. He is the recipient of the Jacobs Fellowship, the Viterbi Fellowship, the Horev Fellowship, and the Seiden Excellence Award. Currently, he is holding the Schmidt Career Advancement Chair in Artificial Intelligence.