Revisiting forward-looking GFlowNets

Detalhes bibliográficos
Autor(a) principal: Silva, Tiago da
Data de Publicação: 2023
Idioma: eng
Título da fonte: Repositório Institucional do FGV (FGV Repositório Digital)
Texto Completo: https://hdl.handle.net/10438/35366
Resumo: As redes gerativas de fluxo (GFlowNets, na sigla em inglês) constituem uma família de algoritmos escaláveis projetados para amostrar de uma distribuição não normalizada com um suporte composicional. Notavelmente, trabalhos recentes exploram a formulação forward-looking de GFlowNets para alcançar desempenho notável em tarefas que formulam a otimização como amostragem, concentrando-se em regiões elevadas. No entanto, este trabalho demonstra que as GFlowNets forward-looking (FL-GFlowNets) violam suposições básicas que fundamentam a corretude da amostragem das GFlowNets. Além disso, projetamos um modelo alternativo que adere aos mesmos princípios das FL-GFlowNets e mostramos que ele gera de maneira comprovada amostras da distribuição correta quando treinado adequadamente. Nossos experimentos validam a teoria delineada e revelam as razões por trás dos resultados aparentemente bons obtidos pelas FL-GFlowNets.
id FGV_5f0be133bc91250ce06621ef7c802e3a
oai_identifier_str oai:repositorio.fgv.br:10438/35366
network_acronym_str FGV
network_name_str Repositório Institucional do FGV (FGV Repositório Digital)
repository_id_str 3974
spelling Silva, Tiago daEscolas::EMApOliveira, DarioSilva, EliezerMesquita, Diego Parente Paiva2024-05-29T18:10:42Z2024-05-29T18:10:42Z2023-12-11https://hdl.handle.net/10438/35366As redes gerativas de fluxo (GFlowNets, na sigla em inglês) constituem uma família de algoritmos escaláveis projetados para amostrar de uma distribuição não normalizada com um suporte composicional. Notavelmente, trabalhos recentes exploram a formulação forward-looking de GFlowNets para alcançar desempenho notável em tarefas que formulam a otimização como amostragem, concentrando-se em regiões elevadas. No entanto, este trabalho demonstra que as GFlowNets forward-looking (FL-GFlowNets) violam suposições básicas que fundamentam a corretude da amostragem das GFlowNets. Além disso, projetamos um modelo alternativo que adere aos mesmos princípios das FL-GFlowNets e mostramos que ele gera de maneira comprovada amostras da distribuição correta quando treinado adequadamente. Nossos experimentos validam a teoria delineada e revelam as razões por trás dos resultados aparentemente bons obtidos pelas FL-GFlowNets.engGFlowNetsMachine LearningMatemáticaRevisiting forward-looking GFlowNetsTCinfo:eu-repo/semantics/publishedVersioninfo:eu-repo/semantics/openAccessreponame:Repositório Institucional do FGV (FGV Repositório Digital)instname:Fundação Getulio Vargas (FGV)instacron:FGVORIGINALmodelo_tcc_emap_main.pdfmodelo_tcc_emap_main.pdfapplication/pdf1007790https://repositorio.fgv.br/bitstreams/632c8016-b07a-4274-994c-fd051bc5898b/download6c90e8a81e1c28f0dfa13baf8f72f1a9MD51LICENSElicense.txtlicense.txttext/plain; charset=utf-85112https://repositorio.fgv.br/bitstreams/7eabc93d-bf0d-48d8-bbd1-25f69ceb2468/download2a4b67231f701c416a809246e7a10077MD52TEXTmodelo_tcc_emap_main.pdf.txtmodelo_tcc_emap_main.pdf.txtExtracted texttext/plain40578https://repositorio.fgv.br/bitstreams/8137e918-6fea-435d-9ab5-978efdc90641/downloadbc80395de3e983e6e860cd361cc7ea07MD53THUMBNAILmodelo_tcc_emap_main.pdf.jpgmodelo_tcc_emap_main.pdf.jpgGenerated Thumbnailimage/jpeg2746https://repositorio.fgv.br/bitstreams/58db8f9a-38c6-477b-999e-355fc7231fee/downloaded647c8c948008d6fd9666155743e4a9MD5410438/353662024-05-29 18:32:13.902open.accessoai:repositorio.fgv.br:10438/35366https://repositorio.fgv.brRepositório InstitucionalPRIhttp://bibliotecadigital.fgv.br/dspace-oai/requestopendoar:39742024-05-29T18:32:13Repositório Institucional do FGV (FGV Repositório Digital) - Fundação Getulio Vargas (FGV)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
dc.title.eng.fl_str_mv Revisiting forward-looking GFlowNets
title Revisiting forward-looking GFlowNets
spellingShingle Revisiting forward-looking GFlowNets
Silva, Tiago da
GFlowNets
Machine Learning
Matemática
title_short Revisiting forward-looking GFlowNets
title_full Revisiting forward-looking GFlowNets
title_fullStr Revisiting forward-looking GFlowNets
title_full_unstemmed Revisiting forward-looking GFlowNets
title_sort Revisiting forward-looking GFlowNets
author Silva, Tiago da
author_facet Silva, Tiago da
author_role author
dc.contributor.unidadefgv.por.fl_str_mv Escolas::EMAp
dc.contributor.member.none.fl_str_mv Oliveira, Dario
Silva, Eliezer
dc.contributor.author.fl_str_mv Silva, Tiago da
dc.contributor.advisor1.fl_str_mv Mesquita, Diego Parente Paiva
contributor_str_mv Mesquita, Diego Parente Paiva
dc.subject.eng.fl_str_mv GFlowNets
Machine Learning
topic GFlowNets
Machine Learning
Matemática
dc.subject.area.por.fl_str_mv Matemática
description As redes gerativas de fluxo (GFlowNets, na sigla em inglês) constituem uma família de algoritmos escaláveis projetados para amostrar de uma distribuição não normalizada com um suporte composicional. Notavelmente, trabalhos recentes exploram a formulação forward-looking de GFlowNets para alcançar desempenho notável em tarefas que formulam a otimização como amostragem, concentrando-se em regiões elevadas. No entanto, este trabalho demonstra que as GFlowNets forward-looking (FL-GFlowNets) violam suposições básicas que fundamentam a corretude da amostragem das GFlowNets. Além disso, projetamos um modelo alternativo que adere aos mesmos princípios das FL-GFlowNets e mostramos que ele gera de maneira comprovada amostras da distribuição correta quando treinado adequadamente. Nossos experimentos validam a teoria delineada e revelam as razões por trás dos resultados aparentemente bons obtidos pelas FL-GFlowNets.
publishDate 2023
dc.date.issued.fl_str_mv 2023-12-11
dc.date.accessioned.fl_str_mv 2024-05-29T18:10:42Z
dc.date.available.fl_str_mv 2024-05-29T18:10:42Z
dc.type.driver.fl_str_mv TC
dc.type.status.fl_str_mv info:eu-repo/semantics/publishedVersion
status_str publishedVersion
dc.identifier.uri.fl_str_mv https://hdl.handle.net/10438/35366
url https://hdl.handle.net/10438/35366
dc.language.iso.fl_str_mv eng
language eng
dc.rights.driver.fl_str_mv info:eu-repo/semantics/openAccess
eu_rights_str_mv openAccess
dc.source.none.fl_str_mv reponame:Repositório Institucional do FGV (FGV Repositório Digital)
instname:Fundação Getulio Vargas (FGV)
instacron:FGV
instname_str Fundação Getulio Vargas (FGV)
instacron_str FGV
institution FGV
reponame_str Repositório Institucional do FGV (FGV Repositório Digital)
collection Repositório Institucional do FGV (FGV Repositório Digital)
bitstream.url.fl_str_mv https://repositorio.fgv.br/bitstreams/632c8016-b07a-4274-994c-fd051bc5898b/download
https://repositorio.fgv.br/bitstreams/7eabc93d-bf0d-48d8-bbd1-25f69ceb2468/download
https://repositorio.fgv.br/bitstreams/8137e918-6fea-435d-9ab5-978efdc90641/download
https://repositorio.fgv.br/bitstreams/58db8f9a-38c6-477b-999e-355fc7231fee/download
bitstream.checksum.fl_str_mv 6c90e8a81e1c28f0dfa13baf8f72f1a9
2a4b67231f701c416a809246e7a10077
bc80395de3e983e6e860cd361cc7ea07
ed647c8c948008d6fd9666155743e4a9
bitstream.checksumAlgorithm.fl_str_mv MD5
MD5
MD5
MD5
repository.name.fl_str_mv Repositório Institucional do FGV (FGV Repositório Digital) - Fundação Getulio Vargas (FGV)
repository.mail.fl_str_mv
_version_ 1827846516274888704

Registros relacionados