Charles Sutton
http://homepages.inf.ed.ac.uk/csutton/
Wednesday 28th November 2012
Time: 4pm
B10 Basement Seminar Room
Alexandra House, 17 Queen Square, London, WC1N 3AR
Quasi-Newton Methods and Continuous Relaxations for MCMC
The performance of Markov chain Monte Carlo (MCMC) methods is often sensitive to the scaling and correlations between the random variables of interest. An important source of information about the local correlation and scale is given by the Hessian matrix of the target distribution, but this is often either computationally expensive or infeasible. I will discuss MCMC methods that make use of quasi-Newton approximations, approximating the Hessian of the target distribution from previous samples and gradients generated by the sampler. A disadvantage of this idea, however, is that in many applications, e.g., in text applications, the variables of interest are discrete rather than continuous, so . So I will also present a new application of an well known identity, called the "Gaussian integral trick", which provides a way to do continuous relaxation for MCMC.
Joint work with Yichuan Zhang, Amos Storkey, and Zoubin Ghahramani
.