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 language | English |
---|---|
Title of host publication | AI 2024: Advances in Artificial Intelligence |
Subtitle of host publication | 37th Australasian Joint Conference on Artificial Intelligence, AI 2024, Proceedings, Part II |
Editors | Mingming Gong, Yiliao Song, Yun Sing Koh, Wei Xiang, Derui Wang |
Place of Publication | Singapore Singapore |
Publisher | Springer |
Pages | 107-136 |
Number of pages | 30 |
ISBN (Electronic) | 9789819603510 |
ISBN (Print) | 9789819603503 |
DOIs | |
Publication status | Published - 2025 |
Externally published | Yes |
Event | Australasian Joint Conference on Artificial Intelligence 2024 - Melbourne, Australia Duration: 25 Nov 2024 → 29 Nov 2024 Conference number: 37th https://ajcai2024.org/ (Conference website) https://doi.org/10.1007/978-981-96-0351-0 (Conference proceedings) |
Publication series
Name | Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) |
---|---|
Volume | 15443 LNAI |
ISSN (Print) | 0302-9743 |
ISSN (Electronic) | 1611-3349 |
Conference
Conference | Australasian Joint Conference on Artificial Intelligence 2024 |
---|---|
Abbreviated title | AJCAI 2024 |
Country/Territory | Australia |
City | Melbourne |
Period | 25/11/24 → 29/11/24 |
Internet address |
|
Keywords
- Bayesian deep learning
- Pathwise gradient
- Reparametrization trick
- Stochastic variational inference
- Zero-variance control variates