Handle URI:
http://hdl.handle.net/10754/624115
Title:
Transport maps and dimension reduction for Bayesian computation
Authors:
Marzouk, Youssef
Abstract:
We introduce a new framework for efficient sampling from complex probability distributions, using a combination of optimal transport maps and the Metropolis-Hastings rule. The core idea is to use continuous transportation to transform typical Metropolis proposal mechanisms (e.g., random walks, Langevin methods) into non-Gaussian proposal distributions that can more effectively explore the target density. Our approach adaptively constructs a lower triangular transport map—an approximation of the Knothe-Rosenblatt rearrangement—using information from previous MCMC states, via the solution of an optimization problem. This optimization problem is convex regardless of the form of the target distribution. It is solved efficiently using a Newton method that requires no gradient information from the target probability distribution; the target distribution is instead represented via samples. Sequential updates enable efficient and parallelizable adaptation of the map even for large numbers of samples. We show that this approach uses inexact or truncated maps to produce an adaptive MCMC algorithm that is ergodic for the exact target distribution. Numerical demonstrations on a range of parameter inference problems show order-of-magnitude speedups over standard MCMC techniques, measured by the number of effectively independent samples produced per target density evaluation and per unit of wallclock time. We will also discuss adaptive methods for the construction of transport maps in high dimensions, where use of a non-adapted basis (e.g., a total order polynomial expansion) can become computationally prohibitive. If only samples of the target distribution, rather than density evaluations, are available, then we can construct high-dimensional transformations by composing sparsely parameterized transport maps with rotations of the parameter space. If evaluations of the target density and its gradients are available, then one can exploit the structure of the variational problem used for map construction. In both settings, we will show links to recent ideas for dimension reduction in inverse problems.
Conference/Event name:
Advances in Uncertainty Quantification Methods, Algorithms and Applications (UQAW 2015)
Issue Date:
7-Jan-2015
Type:
Presentation
Additional Links:
http://mediasite.kaust.edu.sa/Mediasite/Play/15ceede106c84ecf9d2c35042a99f2621d?catalog=ca65101c-a4eb-4057-9444-45f799bd9c52
Appears in Collections:
Conference on Advances in Uncertainty Quantification Methods, Algorithms and Applications (UQAW 2015)

Full metadata record

DC FieldValue Language
dc.contributor.authorMarzouk, Youssefen
dc.date.accessioned2017-06-05T08:35:49Z-
dc.date.available2017-06-05T08:35:49Z-
dc.date.issued2015-01-07-
dc.identifier.urihttp://hdl.handle.net/10754/624115-
dc.description.abstractWe introduce a new framework for efficient sampling from complex probability distributions, using a combination of optimal transport maps and the Metropolis-Hastings rule. The core idea is to use continuous transportation to transform typical Metropolis proposal mechanisms (e.g., random walks, Langevin methods) into non-Gaussian proposal distributions that can more effectively explore the target density. Our approach adaptively constructs a lower triangular transport map—an approximation of the Knothe-Rosenblatt rearrangement—using information from previous MCMC states, via the solution of an optimization problem. This optimization problem is convex regardless of the form of the target distribution. It is solved efficiently using a Newton method that requires no gradient information from the target probability distribution; the target distribution is instead represented via samples. Sequential updates enable efficient and parallelizable adaptation of the map even for large numbers of samples. We show that this approach uses inexact or truncated maps to produce an adaptive MCMC algorithm that is ergodic for the exact target distribution. Numerical demonstrations on a range of parameter inference problems show order-of-magnitude speedups over standard MCMC techniques, measured by the number of effectively independent samples produced per target density evaluation and per unit of wallclock time. We will also discuss adaptive methods for the construction of transport maps in high dimensions, where use of a non-adapted basis (e.g., a total order polynomial expansion) can become computationally prohibitive. If only samples of the target distribution, rather than density evaluations, are available, then we can construct high-dimensional transformations by composing sparsely parameterized transport maps with rotations of the parameter space. If evaluations of the target density and its gradients are available, then one can exploit the structure of the variational problem used for map construction. In both settings, we will show links to recent ideas for dimension reduction in inverse problems.en
dc.relation.urlhttp://mediasite.kaust.edu.sa/Mediasite/Play/15ceede106c84ecf9d2c35042a99f2621d?catalog=ca65101c-a4eb-4057-9444-45f799bd9c52en
dc.titleTransport maps and dimension reduction for Bayesian computationen
dc.typePresentationen
dc.conference.dateJanuary 6-9, 2015en
dc.conference.nameAdvances in Uncertainty Quantification Methods, Algorithms and Applications (UQAW 2015)en
dc.conference.locationKAUSTen
dc.contributor.institutionMassachusetts Institute of Technologyen
All Items in KAUST are protected by copyright, with all rights reserved, unless otherwise indicated.