Statistics Seminar: Matt Wand -- SEMIPARAMETRIC MEAN FIELD VARIATIONAL BAYES

SEMIPARAMETRIC MEAN FIELD VARIATIONAL BAYES

            Professor Matt Wand
         University of Technology, Sydney
   
   A pervasive theme impacting Statistics in the mid-2010s
is the increasing prevalence of data that are big in terms
of volume and/or velocity. Mean field variational Bayes 
methodology is one means of confronting this sea-change.
It delivers fast approximate Bayesian inference in real time. 
Vanilla mean field variational Bayes is inherently 
*nonparametric* in that the functional forms of the 
approximate posterior density functions that it produces 
depend only on the mean field restrictions.
However, these forms can be quite complicated when 
non-conjugacies arise. A possible remedy is *semiparametric*
mean field variational Bayes, in which the complicated 
functional forms are replaced by simpler forms at the outset. 
This leads to new challenges. Some of these will be discussed, 
solved and illustrated for certain non-standard regression models.