Bayesian Statistics Seminar
North Carolina State University

presents

NAME: Dr. Anirban Bhattachary

AFFILIATION: Duke University

Title:  Bayesian Shrinkage

Abstract: Penalized regression methods, such as L1 regularization, are
routinely used in high-dimensional applications, and there is a rich
literature on optimality properties under sparsity assumptions.  In
the Bayesian paradigm, sparsity is routinely induced through
two-component mixture priors having a probability mass at zero, but
such priors encounter daunting computational problems in high
dimensions.  This has motivated an amazing variety of continuous
shrinkage priors, which can be expressed as global-local scale
mixtures of Gaussians, facilitating computation.  In sharp contrast to
the corresponding frequentist literature, very little is known about
the properties of such priors.  Focusing on a broad class of shrinkage
priors, we provide precise results on prior and posterior
concentration.  Interestingly, we demonstrate that most commonly used
shrinkage priors,  including the Bayesian Lasso, are suboptimal in
high-dimensional settings.  A new class of Dirichlet-Laplace (DL)
priors are proposed, which possess optimal concentration and lead to
efficient posterior computation exploiting results from normalized
random measure theory.

ABSTRACT: TBA

Thursday, April, 25, 2013

4:00 - 5:00 pm

1108 SAS Hall

           Back