Sticking the Landing: Simple, Lower-Variance Gradient Estimators for Variational Inference
Abstract
We propose a simple and general variant of the standard reparameterized gradient estimator for the variational evidence lower bound. Specifically, we remove a part of the total derivative with respect to the variational parameters that corresponds to the score function. Removing this term produces an unbiased gradient estimator whose variance approaches zero as the approximate posterior approaches the exact posterior. We analyze the behavior of this gradient estimator theoretically and empirically, and generalize it to more complex variational distributions such as mixtures and importance-weighted posteriors.
Cite
Text
Roeder et al. "Sticking the Landing: Simple, Lower-Variance Gradient Estimators for Variational Inference." Neural Information Processing Systems, 2017.Markdown
[Roeder et al. "Sticking the Landing: Simple, Lower-Variance Gradient Estimators for Variational Inference." Neural Information Processing Systems, 2017.](https://mlanthology.org/neurips/2017/roeder2017neurips-sticking/)BibTeX
@inproceedings{roeder2017neurips-sticking,
title = {{Sticking the Landing: Simple, Lower-Variance Gradient Estimators for Variational Inference}},
author = {Roeder, Geoffrey and Wu, Yuhuai and Duvenaud, David K.},
booktitle = {Neural Information Processing Systems},
year = {2017},
pages = {6925-6934},
url = {https://mlanthology.org/neurips/2017/roeder2017neurips-sticking/}
}