University of Cambridge > > Microsoft Research Cambridge, public talks > Some recent developments in approximate inference: learning and control

Some recent developments in approximate inference: learning and control

Add to your list(s) Download to your calendar using vCal

If you have a question about this talk, please contact Microsoft Research Cambridge Talks Admins.

I’ll discuss two pieces of work on inference in probabilistic models:

The first concerns a very general class of Bayesian Linear Models that are widely used in statistics and machine learning. A great deal of research has been carried out on developing approximate inference techniques for this important class. In particular I’ll discuss methods that bound the model likelihood, which is of interest in parameter learning. The well-known `local’ variational methods lower bound the marginal likelihood. Despite their popularity over the last decade, I’ll discuss our recent result that shows that local methods result in weaker bounds than alternative `mean-field’ variational methods. In addition, I’ll discuss the perhaps surprising result that the mean-field bound is concave and discuss how one may make computationally efficient approximations in large-scale models with many thousands of variables.

Lagrange Duality is being increasingly exploited across machine learning but to date has received comparatively little attention in planning and control. For the second part of the talk I’ll discuss an application of Lagrange Duality in learning Markov Decision Process policies. In particular, I’ll discuss the computationally difficult finite-horizon time-independent policy case, and demonstrate how our method exhibits substantially improved performance compared to policy gradients and more recent `EM’ style procedures.

This talk is part of the Microsoft Research Cambridge, public talks series.

Tell a friend about this talk:

This talk is included in these lists:

Note that ex-directory lists are not shown.


© 2006-2022, University of Cambridge. Contact Us | Help and Documentation | Privacy and Publicity