A1098
Title: Bend to mend: Recalibrating variational Bayes for valid uncertainty quantification
Authors: Jiaming Liu - Rice University (United States)
Meng Li - Rice University (United States) [presenting]
Abstract: Variational Bayes (VB) is widely used for scalable approximate Bayesian inference, but its uncertainty quantification (UQ) often lacks validity, typically producing undercovered credible intervals. The aim is to introduce trustworthy variational Bayes (TVB), a method developed to recalibrate UQ for broad classes of VB procedures. TVB follows a bend-to-mend principle: we intentionally misspecify the likelihood using a fractional posterior indexed by a smoothing parameter omega and then select omega via a conformal calibration framework using sample splitting and bootstrap. This correction yields credible intervals with guaranteed frequentist coverage. Sequential and grid-search strategies are presented for selecting omega, and how TVB enables parallel, parameter-agnostic inference is demonstrated. Results from canonical models and simulations show that TVB outperforms standard VB and achieves reliable UQ in finite samples. A real data example is shared to illustrate its practical utility.
