University of Chicago - Department of Statistics
The term "data augmentation" refers to modeling techniques that introduce unobserved data or latent variables which are often useful for constructing statistical algorithms. These methods have found a multitude of applications in both mode finding (e.g., the EM algorithm) and iterative sampling (e.g., the Data-Augmentation algorithm and, more generally, the Gibbs sampler and MCMC). However, deriving efficient data-augmentation schemes, by efficient schemes we mean schemes that result in simple and fast algorithms, remains a matter of art in that it must be worked out on a case-by-case base. In this talk we introduce two general strategies for searching for efficient data-augmentation.