Graduate Seminar: Iterative Join-Graph Propagation

Abstract

Graphical models are a central research topic in Artificial Intelligence and Machine Learning. We introduce Iterative Join-Graph Propagation (IJGP), a parameterized approximate message-passing scheme. Built upon the principles of bounded inference, notably the mini-clustering algorithm, and Pearl’s belief propagation algorithm (BP), IJGP emerges as a novel approach within the framework of Generalized Belief Propagation algorithms. This framework establishes connections with approximate algorithms from statistical physics, and also facilitates connections to iterative decoding algorithms in information theory. Empirical evaluations demonstrate that IJGP surpasses both mini-clustering and belief propagation, along with several other state-of-the-art algorithms, across various classes of networks. Additionally, we provide insights into the accuracy of iterative BP and IJGP by drawing connections to well-known classes of constraint propagation schemes.

Speaker Biography

Robert Mateescu is a Senior Principal Researcher at Western Digital and an Adjunct Faculty in Computer Science at San Francisco State University, currently teaching Deep Learning. He conducts research in artificial intelligence and machine learning, algorithm design, information theory and distributed systems. He received the B.Sc. in Mathematics and Computer Science from the University of Bucharest, Romania, and the M.S. and Ph.D. degrees in Information and Computer Science from the University of California, Irvine. Dr. Mateescu has extensive academic and industrial research experience, including a postdoc at Caltech, working at Microsoft Research in Cambridge, UK, working in the Research Lab of Hitachi Global Storage Technologies, and in the Research Lab of Western Digital and SanDisk. He is a Senior Member of IEEE, has published over 40 papers and holds 38 US patents. His papers were recognized in journals and conferences, including winning the 2009 IEEE Communications Society best paper award in Signal Processing and Coding for Data Storage, the 2016 IEEE International Conference on Communications best paper award, the 2016-2017 IEEE Data Storage best student paper award, and the 2020 Persistent Impact Prize in the Non-Volatile Memories Workshop. Beyond his professional endeavors, Robert is passionate about the game of Go, has studied as an apprentice professional Go player in Japan, and held the title of European Go Champion.