Corbinian Schlösser (INRIA Paris)

Title : Identifying low dimensional systems from high dimensional observations

Abstract : Motivated by the increasing availability of motion data through multiple sensors and cameras, we address the problem of identifying low-dimensional autonomous systems for which only high-dimensional measurements are observed. We do not assume prior knowledge on the latent dimension or its dynamics. However, we assume that the observation map is injective which implies that the observations follow an autonomous system. This is our starting point for reducing a general and non-linear identification problem towards a (tractable) convex optimization problem. We discuss different obstacles that came across our route, several limiting factors, and situations for which our approach provides feasible solutions to the identification problem. This is joint and ongoing work with Francis Bach and Alessandro Rudi.

The seminar will take place in Room S08 at the Faculty of Sciences.