Speaker:
Affiliation:
Date:
Venue:
Abstract:
Collapsed variational Bayes (CVB) offers a potential approach for improving upon mean field variational Bayes (VB) methods. Both VB and CVB are fast approximate Bayesian inference alternatives to Markov chain Monte Carlo methods. However, VB methods can fail for high-dimensional/low sample size model selection problems. In this talk we show how CVB can successfully solve model selection problems for generalized linear models with high-dimensions and low sample size, overcoming the problems associated with VB for the same model. In our numerical examples we show empirically that CVB approaches are extremely efficient, and perform well in terms of variable selection and prediction in comparison to some popular alternative model selection methods for the same task.
Joint work with Chong You.
