Year: 2013
Type: dissertation
Abstract: Central to many statistical inference problems is the computation ofsome quantities defined over variables that can be fruitfully modeledin terms of graphs. Examples of such quantities include marginaldistributions over graphical models and empirical average ofobservations over sensor networks. For practical purposes, distributedmessage-passing algorithms are well suited to deal with suchproblems. In particular, the computation is broken down into piecesand distributed among different nodes. Following some localcomputations, the intermediate results are shared among neighboringnodes via the so called messages. The process is repeated until thedesired quantity is obtained. These distributed inference algorithmshave two primary aspects: statistical properties, in whichcharacterize how mathematically sound an algorithm is, andcomputational complexity that describes the efficiency of a particularalgorithm. In this thesis, we propose low-complexity (efficient),message-passing algorithms as alternatives to some well knowninference problems while providing rigorous mathematical analysis oftheir performances. These problems include the computation of themarginal distribution via belief propagation for discrete as well ascontinuous random variables, and the computation of the average ofdistributed observations in a noisy sensor network via gossip-typealgorithms.
Author: Nima Noorshams
Institution: University of California, Berkeley
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7058112442
