Stat Cafe - Dr. Pritam Dey
TAVIE: A General Tangent Approximation Framework for Strongly Super-Gaussian Likelihoods
- Time: Tuesday 10/21/2025 from 11:10 AM to 12:25 PM
- Location: BLOC 448
Description
Variational inference (VI) has become a cornerstone of scalable Bayesian computation, yet many approaches rely on model-specific tricks or the “alchemy” of black-box optimization. In this talk, I will introduce TAVIE—a Tangent Approximation Variational InferencE framework that unifies a broad class of strongly super-Gaussian likelihoods under the umbrella of tangent minorization. TAVIE replaces heuristic approximations with a principled framework offering tractable updates and novel theoretical guarantees for algorithmic convergence and non-asymptotic risk bounds. I will outline the algorithmic structure and theoretical foundations, followed by a demonstration of TAVIE’s superior performance through simulations and real data analyses.
Our Speaker
Dr. Pritam Dey is a Postdoctoral Researcher in the Department of Statistics at Texas A&M University. His research focuses on Bayesian methods for scientific machine learning and variational inference, with applications in spatial transcriptomics and multi-omic data integration. He received his Ph.D. in Statistical Science from Duke University.