Stochastic Systems Group
Home Research Group Members Programs  
Demos Calendar Publications Mission Statement Alumni

SSG Seminar Abstract


Multi-message Passing Algorithms for Inference in Loopy Gaussian Graphical Models

Ying Liu
SSG, MIT


Belief propagation is accurate and efficient for tree structured graphs. But the capacity of models on tree is limited. Loopy belief propagation uses an iterative approach to do approximation. However, its convergence and correctness are not guaranteed. Here we aim to establish a message passing algorithm, which includes multiple kinds of messages. The messages have certain header information including the source the messages. Each node processes the incoming messages and sends out new messages according to one of the two updating protocols. In this way, we get accurate inference in Gaussian graphical models with loops. Trade-off between accuracy and efficiency can be made by adjusting the number of the messages based on the complexity of the graphs.



Problems with this site should be emailed to jonesb@mit.edu