Gradients should stay on path: better estimators of the reverse- and forward KL divergence for normalizing flows

dc.contributor.authorVaitl, Lorenz
dc.contributor.authorNicoli, Kim A
dc.contributor.authorNakajima, Shinichi
dc.contributor.authorKessel, Pan
dc.date.accessioned2022-11-21T09:43:37Z
dc.date.available2022-11-21T09:43:37Z
dc.date.issued2022-10-19
dc.date.updated2022-11-10T12:55:17Z
dc.description.abstractWe show how to use the path-wise derivative estimator for both the forward reverse Kullback–Leibler divergence for any practically invertible normalizing flow. The resulting path-gradient estimators are straightforward to implement, have lower variance, and lead not only to faster convergence of training but also to better overall approximation results compared to standard total gradient estimators. We also demonstrate that path-gradient training is less susceptible to mode-collapse. In light of our results, we expect that path-gradient estimators will become the new standard method to train normalizing flows for variational inference.en
dc.description.sponsorshipBMBF, 01IS18025A, Verbundprojekt BIFOLD-BBDC: Berlin Institute for the Foundations of Learning and Data
dc.description.sponsorshipBMBF, 01IS18037A, Verbundprojekt BIFOLD-BZML: Berlin Institute for the Foundations of Learning and Data
dc.identifier.eissn2632-2153
dc.identifier.urihttps://depositonce.tu-berlin.de/handle/11303/17691
dc.identifier.urihttps://doi.org/10.14279/depositonce-16477
dc.language.isoen
dc.rights.urihttp://creativecommons.org/licenses/by/4.0
dc.subject.ddc004 Datenverarbeitung; Informatikde
dc.subject.othervariational inference
dc.subject.otherpath gradients
dc.subject.othermode dropping
dc.subject.othernormalizing flows
dc.subject.otherimportance sampling
dc.titleGradients should stay on path: better estimators of the reverse- and forward KL divergence for normalizing flows
dc.typeArticle
dc.type.versionsubmittedVersion
dcterms.bibliographicCitation.articlenumber045006
dcterms.bibliographicCitation.doi10.1088/2632-2153/ac9455
dcterms.bibliographicCitation.issue4
dcterms.bibliographicCitation.journaltitleMachine Learning: Science and Technologyen
dcterms.bibliographicCitation.originalpublishernameIOP
dcterms.bibliographicCitation.originalpublisherplaceBristol
dcterms.bibliographicCitation.volume3
tub.accessrights.dnbfree
tub.affiliationFak. 4 Elektrotechnik und Informatik::Inst. Softwaretechnik und Theoretische Informatik::FG Maschinelles Lernen
tub.publisher.universityorinstitutionTechnische Universität Berlin

Files

Original bundle
Now showing 1 - 1 of 1
Loading…
Thumbnail Image
Name:
mlst_3_4_045006.pdf
Size:
3.55 MB
Format:
Adobe Portable Document Format
License bundle
Now showing 1 - 1 of 1
No Thumbnail Available
Name:
license.txt
Size:
4.86 KB
Format:
Item-specific license agreed upon to submission
Description:

Collections