Course Description

Location has changed to  room R42.4.110.   See http://www.ulb.ac.be/campus/solbosch/plan-R42.html

L'Ecole polytechnique de Bruxelles together with MUSICS doctoral school are pleased to announce the following IEEE Distinguished Lectures of Prof. James V. Candy, University of California in Santa Barbara on

* Bayesian Model-based Signal Processing: Particle Filtering Techniques (Theory)

Bayesian signal processing is primarily concerned with the estimation of the underlying posterior distribution governing the problem. Once extracted, all of the required statistical information contained in the stochastic process along with its meaningful statistics is at hand for analysis. The Bayesian approach is the “next step” in model-based signal processing in that it not only seeks probabilistic distributions to represent the underlying stochastic processes but also enables an entire framework that allows the incorporation of physics-based models, measurements and noise processes into the processing scheme.
The development of Bayesian processors is discussed after developing the basic theory from a Bayesian perspective. Starting with the Bayesian construct and then developing the underlying framework for problem solving, the idea of simulation-based sampling theory leading to sequential Bayesian processing is developed. Next the fundamental Bayesian processor currently applied in the signal processing literature and is discussed along with some of its variants. Simple applications and examples are used throughout to solidify these ideas in a meaningful manner.

* Bayesian Model-based Signal Processing: Particle Filtering Techniques (Advanced Processors and Applications)

Bayesian signal processing is a novel area migrated from statistical sampling theory enabling the solution of problems in which the posterior distribution is multimodal (many peaks) as compared to the usual single mode solutions (Gaussian distributions). The introduction of sequential Bayesian processors rediscovered in the early 1990’s has had a huge impact in signal processing enabling the solution of highly complex problems.
In this lecture, we concentrate on sequential Bayesian model-based processors and discuss a variety of the basic structures that are available in the literature. These processors are termed particle filters. More properly, a particle filter is a sequential Monte Carlo technique capable of solving nonlinear, multimodal problems by estimating the posterior probability distribution enabling the extraction of a wide variety of signal estimators (conditional mean, maximum a-posteriori, etc.) from noisy measurement data. We show how these powerful processors can incorporate the underlying physics, measurement instrumentation and noise into a model-based signal processing scheme to extract the desired information from the data. Thus, we develop the particle filter within the “model-based” framework to provide the Bayesian model-based solution applicable to a large variety of applications.
After a brief review of importance sampling and recalling the sequential Bayesian construct, we discuss the minimum variance particle filter which is difficult to implement due to its dependence on past data. Next the evolution of perhaps the most popular and simple design called the “bootstrap” particle filter is reviewed. The development of more sophisticated particle filter designs including the regularized, Markov chain Monte Carlo (MCMC) and the linearized particle filters are developed. Detailed applications including oceanic processing, trajectory estimation/identification, nuclear physics and parametrically adaptive processing are discussed to elucidate some of these ideas and give a practical feel for performance.

When ? May 30th 2013

Where ? ULB LISA seminar room on Solbosch campus

Bâtiment L, Porte E, 3ème niveau
Salle L3-219.
http://www.ulb.ac.be/campus/solbosch/plan-en-L-en.html

The schedule is:
Lecture #1: 10:15 to 12:00
Lecture #2: 13:30 to 15:00

Lunch meals can be taken on Solbosch campus.

Registration free but mandatory on MUSICS website until May 23rd 2013

Local coordinator : Jean-Pierre Hermand (jhermand@ulb.ac.be)