Boston University
Enhancing the reliability of general-purpose algorithms for approximate Bayesian inference
Abstract
dc:description.abstractRecent advances in computational statistics and machine learning have spurred the development of scalable Bayesian inference methods capable of handling increasingly complex models and large datasets. For example, posterior approximation algorithms such as stochastic gradient Langevin dynamics (SGLD) and variational inference offer both flexibility and computational efficiency. While essential for making reliable approximation inference, robust validation of approximation quality and accurate uncertainty quantification remain underdeveloped. This thesis addresses these challenges by introducing novel theoretical insights and practical methodologies for Bayesian inference. First, we extend the a priori analysis of algorithm accuracy by offering a unified framework for evaluating convergence accuracy, key properties, and easily checkable criteria for a broad class of target distributions satisfying strong log-concavity and the Euclidean log-Sobolev inequality. To evaluate the quality of variational approximations, we propose a robust post hoc diagnostic tool for variational approximations, the TArgeted Diagnostic for Distribution Approximation Accuracy (TADDAA), which utilizes multiple short parallel Markov chain Monte Carlo (MCMC) runs to provide lower bounds on the error of posterior quantities. To better quantify uncertainty and enhance the efficiency of stochastic optimization algorithms like Stochastic Gradient Descent (SGD) and subsampling MCMC algorithms like SGLD, we propose a new gradient linearization approach to accurately estimate the stationary covariance at fixed non-vanishing learning rates. All findings are extensively validated through experiments conducted on both synthetic and real-world 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
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Wang, Yu
- Advisor dc:contributor.advisor
-
- Huggins, Jonathan H.
Subjects
dc:subject × 1Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/2144/53264