MFDTs: Mean field dynamic trees

NJ Adams*, AJ Storkey, Z Ghahramani, CKI Williams

*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

Tree structured belief networks al-e attractive for image segmentation tasks. However; networks with fixed architectures are not view suitable as they lead to blocky artefacts, and led to the introduction of Dynamic Trees (DTs) in [6]. The Dynamic Trees architecture provides a prior distribution over tree structures, and in [6] simulated annealing (SA) was used to search for structures with high pasterior probability. In this paper, Mle introduce a mean field approach to inference in DTs. We find that the mean field method captures the posterior better than just using the maximum a posterior solution found by SA.

Original languageEnglish
Title of host publication15TH INTERNATIONAL CONFERENCE ON PATTERN RECOGNITION, VOL 3, PROCEEDINGS
EditorsA Sanfeliu, JJ Villanueva, M Vanrell, R Alquezar, T Huang, J Serra
Place of PublicationLOS ALAMITOS
PublisherInstitute of Electrical and Electronics Engineers (IEEE)
Pages147-150
Number of pages4
ISBN (Print)0-7695-0751-4
DOIs
Publication statusPublished - 2000
Event15th International Conference on Pattern Recognition (ICPR-2000) - BARCELONA, Spain
Duration: 3 Sep 20007 Sep 2000

Publication series

NameINTERNATIONAL CONFERENCE ON PATTERN RECOGNITION
PublisherIEEE COMPUTER SOC
ISSN (Print)1051-4651

Conference

Conference15th International Conference on Pattern Recognition (ICPR-2000)
Country/TerritorySpain
Period3/09/007/09/00

Fingerprint

Dive into the research topics of 'MFDTs: Mean field dynamic trees'. Together they form a unique fingerprint.

Cite this