Dual-decomposition (DD) methods are quickly becoming important tools for estimating the minimum energy state of a graphical model. DD methods decompose a complex model into a collection of simpler subproblems that can be solved exactly (such as trees), that in combination provide upper and lower bounds on the exact solution. Subproblem choice can play a major role: larger subproblems tend to improve the bound more per iteration, while smaller subproblems enable highly parallel solvers and can benefit from re-using past solutions when there are few changes between iterations. We propose an algorithm that can balance many of these aspects to speed up convergence. Our method uses a cluster tree data structure that has been proposed for adaptive exact inference tasks, and we apply it in this paper to dual-decomposition approximate inference. This approach allows us to process large subproblems to improve the bounds at each iteration, while allowing a high degree of parallelizability and taking advantage of subproblems with sparse updates. For both synthetic inputs and a real-world stereo matching problem, we demonstrate that our algorithm is able to achieve significant improvement in convergence time.

Motivated by stochastic systems in which observed evidence and conditional de- pendencies between states of the network change over time, and certain quantities of interest (marginal distributions, likelihood estimates etc.) must be updated, we study the problem of adaptive inference in tree-structured Bayesian networks. We describe an algorithm for adaptive inference that handles a broad range of changes to the network and is able to maintain marginal distributions, MAP estimates, and data likelihoods in all expected logarithmic time. We give an implementation of our algorithm and provide experiments that show that the algorithm can yield up to two orders of magnitude speedups on answering queries and responding to dy- namic changes over the sum-product algorithm.