Pathwise Gradient Variance Reduction with Control Variates in Variational Inference

Kenyon Ng, Susan Wei

Research output: Chapter in Book/Report/Conference proceedingConference PaperOtherpeer-review

Abstract

Variational inference in Bayesian deep learning often involves computing the gradient of an expectation that lacks a closed-form solution. In these cases, pathwise and score-function gradient estimators are the most common approaches. The pathwise estimator is often favoured for its substantially lower variance compared to the score-function estimator, which typically requires variance reduction techniques. However, recent research suggests that even pathwise gradient estimators could benefit from variance reduction. In this work, we review existing control-variates-based variance reduction methods for pathwise gradient estimators to assess their effectiveness. Notably, these methods often rely on integrand approximations and are applicable only to simple variational families. To address this limitation, we propose applying zero-variance control variates to pathwise gradient estimators. This approach offers the advantage of requiring minimal assumptions about the variational distribution, other than being able to sample from it.

Original languageEnglish
Title of host publicationAI 2024: Advances in Artificial Intelligence
Subtitle of host publication37th Australasian Joint Conference on Artificial Intelligence, AI 2024, Proceedings, Part II
EditorsMingming Gong, Yiliao Song, Yun Sing Koh, Wei Xiang, Derui Wang
Place of PublicationSingapore Singapore
PublisherSpringer
Pages107-136
Number of pages30
ISBN (Electronic)9789819603510
ISBN (Print)9789819603503
DOIs
Publication statusPublished - 2025
Externally publishedYes
EventAustralasian Joint Conference on Artificial Intelligence 2024 - Melbourne, Australia
Duration: 25 Nov 202429 Nov 2024
Conference number: 37th
https://ajcai2024.org/ (Conference website)
https://doi.org/10.1007/978-981-96-0351-0 (Conference proceedings)

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume15443 LNAI
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

ConferenceAustralasian Joint Conference on Artificial Intelligence 2024
Abbreviated titleAJCAI 2024
Country/TerritoryAustralia
CityMelbourne
Period25/11/2429/11/24
Internet address

Keywords

  • Bayesian deep learning
  • Pathwise gradient
  • Reparametrization trick
  • Stochastic variational inference
  • Zero-variance control variates

Cite this