| PhD Seminar

Name of the Speaker: Ms. Shivani Bathla (EE17D037)
Guide: Prof. Vinita Vasudevan
Venue: ESB-234 (Malaviya Hall)
Date/Time: 19th January 2024 (Friday), 11:00 AM
Title: Incremental Build-Infer-Approximate (IBIA): A novel framework for approximate inference in probabilistic graphical models

Abstract

Exact inference in probabilistic graphical models (PGM) is intractable and has an exponential dependence on the size of the largest clique in the corresponding clique tree (CT), necessitating approximations. Existing methods for approximate inference either use sampling based methods or use iterative message-passing algorithms, which are slow to converge for many benchmarks. We propose an alternative approach for approximate inference based on an incremental build-infer-approximate (IBIA) paradigm which converts the PGM into a sequence of clique tree forests (SCTF), each with bounded clique sizes. We propose algorithms for the inference of two fundamental queries namely, partition function and posterior marginals, using the SCTF generated by IBIA. In this seminar, I will present the results obtained using the proposed algorithm for inference of partition function and posterior marginals using the IBIA framework. The framework was evaluated using several benchmark sets from recent UAI competitions. It was also used for the estimation of output error rate and circuit error rate in large digital circuits. The results show that our method gives either better or comparable accuracy than the state-of-the-art methods, with competitive runtimes.