Variational Bayesian Approximation for Linear Inverse Problems with a hierarchical prior models
Student No.:50
Time:16:00-17:00, Dec. 9(Mon.)
Instructor:Ali Mohammad-Djafari  [Université Paris Sud]
Place:Conference Room 4, floor 2, Jin Chun Yuan West Building
Starting Date:2013-12-9
Ending Date:2013-12-9




Variational Bayesian Approximation (VBA) method is a recent tool for effective full Bayesian computations. The main idea is to approximate a joint posterior probability law with a simpler,  say a separable probability law to be able to do Bayesian computation more easily.

This approximate solution is obtained by minimizing the Kullback-Leibler divergence  between this separable approximate probability law and the original one.

In this talk, first I explain the basics of this method and then the different steps needed to use it for linear inverse problems where the prior models include hidden variables (Hierarchical prior models) and where the estimation of the hyper parameters has also to be addressed. In particular one specific prior model (Student-t)
is considered and used via a hierarchical representation with hidden variables and the details of the resulted VBA algorithms are given.

Contact person: Jian Sun