Abstract
dc:description.abstractScalable Bayesian inference widely relies on approximate inference algorithms such as Markov chain Monte Carlo (MCMC) and black-box variational inference (BBVI). While MCMC algorithms provide asymptotically exact samples, BBVI, which uses stochastic optimization, yields fast and flexible approximations. However, these modern approximate inference algorithms often lack (1) robust, automated procedures to ensure the reliability of stochastic optimization and (2) quantitative tools to diagnose convergence to stationarity. To address the first challenge, we propose robust and automated black-box variational inference (RABVI), a framework for improving the reliability of BBVI optimization. RABVI is based on rigorously justified automation techniques, uses only a small number of intuitive tuning parameters, and detects inaccurate estimates of the optimal variational approximation. RABVI adaptively decreases the learning rate by detecting convergence of the fixed-learning-rate iterates, then estimates the symmetrized Kullback-Leibler (KL) divergence between the current variational approximation and the optimal one. It also employs a novel optimization termination criterion that allows the user to balance desired accuracy against computational cost. To address the second challenge, we introduce sliced-Wasserstein automated stationarity detection (SWASD), a convergence diagnostic for Markov chains that measures the distributional convergence of the iterates. SWASD uses a block-based approximation to estimate the distance between the marginal distribution of recent Markov chain iterates and the stationary distribution. This approach enables users to estimate when convergence has occurred and how close the iterates are to stationarity. We validate our findings through carefully designed experiments on both synthetic and real-world models and datasets.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Mathematics & Statistics
- Grantor
- Boston University
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Welandawe, Manushi Kushinthara Viduneth
- Advisor dc:contributor.advisor
-
- Huggins, Jonathan H
- ORCID dc:identifier.orcid
Subjects
dc:subject × 7Rights
dc:rights- Statement dc:rights
-
- Attribution 4.0 International
- Licence dc:rights.uri
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/2144/53683