This summer I was lucky enough to collaborate with Lauren Hund from Sandia National Laboratories on some work related to uncertainty quantification (UQ) and Bayesian model calibration. We worked together to extend ideas on power likelihood methods to a model calibration problem originally published earler this year in JRSS C. Novin Ghaffari, another PhD student at UT, also worked with Lauren, though Novin worked on a problem of combining subset posteriors and so his work did not intersect much with mine.
All this work culminated in a mini UQ workshop hosted here at UT, and organized by our professor Peter Mueller. Some of Lauren’s coworkers from Sandia attended, as well as Brian Williams from Los Alamos National Laboratories, and Mengyang Gu from Johns Hopkins University. I gave the first talk of the workshop to introduce the problem and discuss the work I had done. The main idea is to temper a misspecified likelihood to achieve nominal coverage of posterior credible intervals. Novin spoke right after me, and there were talks for the rest of day, as well as a lot of lively discussion.
This was my first time working on UQ problems, and I am grateful to both Lauren and Peter for the experience. UQ is not an area of much active research at SDS here at UT (as far as I know), so this was a very exciting and eye-opening experience.
Slides from my talk are available here. I hope to work further with Lauren to get a manuscript of our work on the arxiv, and eventually submit our work for publication in a UQ journal.