News‎ > ‎

Seminar 31/5: Sequential Monte Carlo Mini-Symposium

posted May 22, 2017, 2:12 AM by Allison Hsiang
On Wednesday, May 31, we will be hosting a mini-symposium on Sequential Monte Carlo (SMC) at the museum. We will have two speakers, Fredrik Lindsten and Lawrence Murray, both from the Department of Information Technology at Uppsala University. The talks will take place in the morning, followed by lunch and afternoon discussion for all interested parties.

Welcome!


Title: Divide-and-Conquer Sequential Monte Carlo for Inference in Probabilistic Graphical Models
Speaker: Fredrik Lindsten
Date: May 31, 2017
Time: 10:00-11:00
Place: Naturhistoriska riksmuseet, Vintergatan conference room, 5th floor (near Cosmonova)

Abstract: Probabilistic graphical models (PGMs) are widely used to represent and to reason about underlying structure in high-dimensional probability distributions. We develop a framework for using sequential Monte Carlo
(SMC) methods for inference and learning in general PGMs. Structural information from the PGM is used to decompose the graph into a collection of subgraphs that can be organized in a tree. Based on this we develop a
new class of SMC samplers, Divide-and-Conquer SMC, for performing inference over the tree. We will see how this method extends the standard chain-based SMC framework to a method that naturally runs on trees. We
illustrate empirically that these approaches can outperform standard methods in terms of estimation accuracy. They also open up novel parallel implementation options and the possibility of concentrating the computational effort on the most challenging parts of the problem at hand.

--------------------------------------------------------------------------------------------------

Title: Software and Sequential Monte Carlo
Speaker: Lawrence Murray
Date: May 31, 2017
Time: 11:00-12:00
Place: Naturhistoriska riksmuseet, Vintergatan conference room, 5th floor (near Cosmonova)

Abstract: I will give a brief introduction to two software projects: LibBi and Birch. LibBi is used for state-space modelling on parallel and distributed computing hardware, such as multicore CPUs, GPUs and clusters, using Sequential Monte Carlo (SMC) and particle Markov chain Monte Carlo (PMCMC) methods. It is particularly effective for models with complex nonlinear dynamics, including continuous-time dynamics. Birch is its newer incarnation, currently in development. The idea of Birch is to broaden the class of models and inference methods that can be supported, beyond state-space models and SMC, moving from a model specification language to a general-purpose probabilistic programming language. I will focus in particular on the ability of Birch to solve analytically-tractable substructure in complex models.

--------------------------------------------------------------------------------------------------

Afternoon Discussion Session
Date: May 31, 2017
Time: 13:00-16:00
Place: Naturhistoriska riksmuseet, Sirius conference room, 5th floor (near Cosmonova)
Comments