Belief propagation on factor graph neural networks

Probabilistic graphical models are a statistical framework for conditionally dependent random variables with dependencies represented by graphs. A traditional method to perform inference over these random variables is Belief Propagation. Belief Propagation can be used to compute an exact solution for non-loopy factor graphs. However, when applied to loopy factor graphs, it only estimates marginal probabilities approximately. In this thesis, we propose a Graph Neural Networks (GNN) approach for belief propagation based on message passing mechanisms. In the proposed approach, representations and other functions are learned by the GNN. We apply this approach to the inference of loopy factor graphs. Furthermore, we show that learned representations and functions can also be generalized to factor graphs with different sizes and structures. The results show that our proposal has promising performance compared to the state of the art.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Mechanical Engineering

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2021

Author dc:creator

• Yin, Jialong

Contributors dc:contributor

• Koyejo, Sanmi

• Belief Propagation
• Graph Neural Networks

Statement dc:rights

• Copyright 2020 Jialong Yin

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/109620

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/109620