Sequential Monte Carlo for
Undirected Models
by Firas Hamze
I will be discussing my work on sequential Monte Carlo (SMC) for undirected
graphical models. The objects are (1) to approximate integrals of functions of
the variables, of which the marginal probabilities used in inference are a
special case, and (2) to approximate the normalization constant of the model.
The talk will first introduce the models, including the familiar Markov Random
Fields used in computer vision, and then discuss a more general approach to SMC
than the one we may be accustomed to seeing in the context of particle
filtering. I will then summarize my sequential edge-coupling algorithm and its
relation to annealing-based methods, and its distinction from Markov Chain Monte
Carlo (MCMC).