Gradient estimators for normalizing flows

2024
journal article
article
dc.abstract.enRecently, a machine learning approach to Monte-Carlo simulations called Neural Markov Chain Monte Carlo (NMCMC) is gaining traction. In its most popular form, it uses neural networks to construct normalizing flows which are then trained to approximate the desired target distribution. In this contribution, we present a new gradient estimator for the Stochastic Gradient Descent algorithm (and the corresponding PyTorch implementation) and show that it leads to better training results for the model. For this model, our estimator achieves the same precision in approximately half of the time needed in the standard approach and ultimately provides better estimates of the free energy. We attribute this effect to the lower variance of the new estimator. In contrary to the standard learning algorithm, our approach does not require estimation of the action gradient with respect to the fields, thus has the potential of further speeding up the training for models with more complicated actions.
dc.affiliationWydział Fizyki, Astronomii i Informatyki Stosowanej : Instytut Informatyki Stosowanej
dc.affiliationWydział Fizyki, Astronomii i Informatyki Stosowanej : Instytut Fizyki Teoretycznej
dc.contributor.authorBiałas, Piotr - 127296
dc.contributor.authorKorcyl, Piotr - 125645
dc.contributor.authorStebel, Tomasz - 150140
dc.date.accessioned2024-04-19T10:43:13Z
dc.date.available2024-04-19T10:43:13Z
dc.date.issued2024
dc.date.openaccess0
dc.description.accesstimew momencie opublikowania
dc.description.number3
dc.description.versionostateczna wersja wydawcy
dc.description.volume55
dc.identifier.articleidA2
dc.identifier.doi10.5506/APhysPolB.55.3-A2
dc.identifier.issn0587-4254
dc.identifier.issn1509-5770
dc.identifier.urihttps://ruj.uj.edu.pl/handle/item/330601
dc.languageeng
dc.language.containereng
dc.rightsUdzielam licencji. Uznanie autorstwa 4.0 Międzynarodowa
dc.rights.licenceCC-BY
dc.rights.urihttp://creativecommons.org/licenses/by/4.0/legalcode.pl
dc.share.typeotwarte czasopismo
dc.subtypeArticle
dc.titleGradient estimators for normalizing flows
dc.title.journalActa Physica Polonica, Series B.
dc.typeJournalArticle
dspace.entity.typePublicationen
dc.abstract.en
Recently, a machine learning approach to Monte-Carlo simulations called Neural Markov Chain Monte Carlo (NMCMC) is gaining traction. In its most popular form, it uses neural networks to construct normalizing flows which are then trained to approximate the desired target distribution. In this contribution, we present a new gradient estimator for the Stochastic Gradient Descent algorithm (and the corresponding PyTorch implementation) and show that it leads to better training results for the model. For this model, our estimator achieves the same precision in approximately half of the time needed in the standard approach and ultimately provides better estimates of the free energy. We attribute this effect to the lower variance of the new estimator. In contrary to the standard learning algorithm, our approach does not require estimation of the action gradient with respect to the fields, thus has the potential of further speeding up the training for models with more complicated actions.
dc.affiliation
Wydział Fizyki, Astronomii i Informatyki Stosowanej : Instytut Informatyki Stosowanej
dc.affiliation
Wydział Fizyki, Astronomii i Informatyki Stosowanej : Instytut Fizyki Teoretycznej
dc.contributor.author
Białas, Piotr - 127296
dc.contributor.author
Korcyl, Piotr - 125645
dc.contributor.author
Stebel, Tomasz - 150140
dc.date.accessioned
2024-04-19T10:43:13Z
dc.date.available
2024-04-19T10:43:13Z
dc.date.issued
2024
dc.date.openaccess
0
dc.description.accesstime
w momencie opublikowania
dc.description.number
3
dc.description.version
ostateczna wersja wydawcy
dc.description.volume
55
dc.identifier.articleid
A2
dc.identifier.doi
10.5506/APhysPolB.55.3-A2
dc.identifier.issn
0587-4254
dc.identifier.issn
1509-5770
dc.identifier.uri
https://ruj.uj.edu.pl/handle/item/330601
dc.language
eng
dc.language.container
eng
dc.rights
Udzielam licencji. Uznanie autorstwa 4.0 Międzynarodowa
dc.rights.licence
CC-BY
dc.rights.uri
http://creativecommons.org/licenses/by/4.0/legalcode.pl
dc.share.type
otwarte czasopismo
dc.subtype
Article
dc.title
Gradient estimators for normalizing flows
dc.title.journal
Acta Physica Polonica, Series B.
dc.type
JournalArticle
dspace.entity.typeen
Publication
Affiliations

* The migration of download and view statistics prior to the date of April 8, 2024 is in progress.

Views
15
Views per month
Views per city
Krakow
4
Cieszyn
1
Downloads
bialas_korcyl_stebel_gradient_estimators_for_normalizing_2024.pdf
4