Abstract

want to estimate the path of camera rotation from a video sequence it has taken? Find it hard to probabilistically model rotation matrices and image observations using traditional graphical model? I am going to talk about an alternative representation for distributions, formally called Hilbert space embedding of distribution, which makes this problem easier. In particular, I will focus on extending the Hilbert space embedding approach to handle conditional distributions. I am going to show a simple kernel estimate for the conditional embedding, and explain its connection to ordinary embeddings, kernel dependency estimation and Gaussian process for regression. Conditional embeddings largely extend our ability to manipulate distributions in Hilbert spaces, and as an example, we derive a nonparametric method for modeling dynamical systems where the belief state of the system is maintained as a conditional embedding. This new method is very general in terms of both the domains and the types of distributions that it can handle, and we demonstrate the effectiveness of our method using several dynamical systems. Conditional embeddings are expected to have wider applications beyond modeling dynamical systems.

Venue, Date, and Time

Venue: NSH 1507

Date: Monday, Sept 21, 2009

Time: 12:00 noon