Cold Analysis of Rao-Blackwellized Straight-Through Gumbel-Softmax Gradient Estimator

Kód výsledku v IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F68407700%3A21230%2F23%3A00371050" target="_blank" >RIV/68407700:21230/23:00371050 - isvavai.cz</a>

Výsledek na webu

<a href="https://proceedings.mlr.press/v202/shekhovtsov23a.html" target="_blank" >https://proceedings.mlr.press/v202/shekhovtsov23a.html</a>

DOI - Digital Object Identifier

—

Jazyk výsledku

angličtina

Název v původním jazyce

Cold Analysis of Rao-Blackwellized Straight-Through Gumbel-Softmax Gradient Estimator

Popis výsledku v původním jazyce

Many problems in machine learning require an estimate of the gradient of an expectation in discrete random variables with respect to the sampling distribution. This work is motivated by the development of the Gumbel-Softmax family of estimators, which use a temperature-controlled relaxation of discrete variables. The state-of-the art in this family, the Gumbel-Rao estimator uses an extra internal sampling to reduce the variance, which may be costly. We analyze this estimator and show that it possesses a zero temperature limit with a surprisingly simple closed form. The limit estimator, called ZGR, has favorable bias and variance properties, it is easy to implement and computationally inexpensive. It decomposes as the average of the straight through (ST) estimator and DARN estimator — two basic but not very well performing on their own estimators. We demonstrate that the simple ST–ZGR family of estimators practically dominates in the bias-variance tradeoffs the whole GR family while also outperforming SOTA unbiased estimators.

Název v anglickém jazyce

Cold Analysis of Rao-Blackwellized Straight-Through Gumbel-Softmax Gradient Estimator

Popis výsledku anglicky

Many problems in machine learning require an estimate of the gradient of an expectation in discrete random variables with respect to the sampling distribution. This work is motivated by the development of the Gumbel-Softmax family of estimators, which use a temperature-controlled relaxation of discrete variables. The state-of-the art in this family, the Gumbel-Rao estimator uses an extra internal sampling to reduce the variance, which may be costly. We analyze this estimator and show that it possesses a zero temperature limit with a surprisingly simple closed form. The limit estimator, called ZGR, has favorable bias and variance properties, it is easy to implement and computationally inexpensive. It decomposes as the average of the straight through (ST) estimator and DARN estimator — two basic but not very well performing on their own estimators. We demonstrate that the simple ST–ZGR family of estimators practically dominates in the bias-variance tradeoffs the whole GR family while also outperforming SOTA unbiased estimators.

Druh

D - Stať ve sborníku

CEP obor

—

OECD FORD obor

10201 - Computer sciences, information science, bioinformathics (hardware development to be 2.2, social aspect to be 5.8)

Projekt

<a href="/cs/project/EF16_019%2F0000765" target="_blank" >EF16_019/0000765: Výzkumné centrum informatiky</a><br>

Návaznosti

P - Projekt vyzkumu a vyvoje financovany z verejnych zdroju (s odkazem do CEP)

Rok uplatnění

2023

Kód důvěrnosti údajů

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů

Název statě ve sborníku

International Conference on Machine Learning (ICML)

ISBN

—

ISSN

2640-3498

e-ISSN

2640-3498

Počet stran výsledku

25

Strana od-do

30931-30955

Název nakladatele

Proceedings of Machine Learning Research

Místo vydání

—

Místo konání akce

Honolulu, USA

Datum konání akce

23. 7. 2023

Typ akce podle státní příslušnosti

WRD - Celosvětová akce

Kód UT WoS článku

—