Skip to content. Skip to navigation
CIM Menus
 

Diverse Particle Selection for Inference in Continuous Graphical Models


Jason Pacheco


May 25, 2016 at  11:00 AM
McConnell Engineering Room 437

Abstract:

In this talk I will introduce a family of algorithms for maximum a posteriori (MAP) inference in models of high dimensional continuous random variables. The approach builds on the well-known max-product variant of belief propagation (BP), which provides a computationally efficient alternative to Markov chain Monte Carlo (MCMC) sampling. Motivated by similar ideas in sum-product BP we eliminate restrictions to Gaussian random variables via a particle-based approximation of the continuous max-product messages. Our nonparametric approximation applies to any model for which the probability density can be evaluated in a black-box manner, even for models with no analytic form. Unique to the MAP setting is a need for diversity among hypotheses, to avoid classic particle degeneracies. Using an integer programming (IP) formulation we enforce particle diversity, from which we can recover a set of distinct local maxima. Furthermore, we show that our particle selection is a submodular maximization, thereby allowing efficient greedy selection with a bound on optimality. The approach is validated on a set of distinct problems including human pose estimation from single images and videos as well as estimating protein structure.

Brief Presentation:

Dr. Jason Pacheco obtained his PhD degree in Computer Science at Brown University in Providence, RI, USA. His research focus is in Machine Learning and in particular in approximate Bayesian inference and graphical models. His work aims to create more flexible variational inference algorithms that can be applied to many challenging models in various domains such as computer vision, computational biology, and signal processing.