Partition Tree Weighting
Joel Veness (University of Alberta)
COMPUTER SCIENCE SEMINARDATE: 2012-09-05
TIME: 11:00:00 - 12:00:00
LOCATION: RSISE Seminar Room, ground floor, building 115, cnr. North and Daley Roads, ANU
CONTACT: JavaScript must be enabled to display this email address.
ABSTRACT:
This talk will describe Partition Tree Weighting (PTW), a low-complexity technique for performing exact Bayesian model averaging over a large class of possible temporal segmentations of a stream of sequential data. Its computational efficiency derives from a carefully chosen prior that is closely related to Willems' celebrated Context Tree Weighting algorithm. This choice of prior is justified by a number of theoretical results which suggest that the technique can be robustly applied to better handle many kinds of non-stationary data. To give some examples, PTW is applied to the KT estimator to give rise to a universal algorithm for coding piecewise stationary memoryless sources and an alternate algorithm for tracking. A collection of empirical results will also be presented, including a result that shows PTW can be used to improve the performance of a recently introduced universal data compression algorithm. The talk should be of interest to researchers familiar with information theory, Bayesian methods, online learning, and/or time series methods.
BIO:
Joel Veness did his PhD with Marcus Hutter on Monte-Carlo AIXI approximations and has just completed a two year post-doc at University of Alberta.
