Monte Carlo Variational Auto-Encoders

被引：0

作者：

Thin, Achille ^{[1
]}

Kotelevskii, Nikita ^{[2
]}

Durmus, Alain ^{[3
]}

Panov, Maxim ^{[2
]}

Moulines, Eric ^{[1
,4
]}

Doucet, Arnaud ^{[5
]}

机构：

[1] Univ Paris Saclay, Ecole Polytech, CMAP, Gif Sur Yvette, France

[2] Skolkovo Inst Sci & Technol, CDISE, Moscow, Russia

[3] Ecole Natl Super Paris Saclay, Gif Sur Yvette, France

[4] HSE Univ, HDI Lab, Moscow, Russia

[5] Univ Oxford, Oxford, England

来源：

INTERNATIONAL CONFERENCE ON MACHINE LEARNING, VOL 139 | 2021年 / 139卷

基金：

英国工程与自然科学研究理事会; 俄罗斯科学基金会;

关键词：

D O I：

暂无

中图分类号：

TP18 [人工智能理论];

学科分类号：

081104 ; 0812 ; 0835 ; 1405 ;

摘要：

Variational auto-encoders (VAE) are popular deep latent variable models which are trained by maximizing an Evidence Lower Bound (ELBO). To obtain tighter ELBO and hence better variational approximations, it has been proposed to use importance sampling to get a lower variance estimate of the evidence. However, importance sampling is known to perform poorly in high dimensions. While it has been suggested many times in the literature to use more sophisticated algorithms such as Annealed Importance Sampling (AIS) and its Sequential Importance Sampling (SIS) extensions, the potential benefits brought by these advanced techniques have never been realized for VAE: the AIS estimate cannot be easily differentiated, while SIS requires the specification of carefully chosen backward Markov kernels. In this paper, we address both issues and demonstrate the performance of the resulting Monte Carlo VAEs on a variety of applications.

引用

页码：7258 / 7267

页数：10