An Analysis of Techniques for Building Generative Adversarial Networks

Detalhes bibliográficos
Autor(a) principal: Hoeckler, Patrick
Data de Publicação: 2021
Tipo de documento: Trabalho de conclusão de curso
Idioma: eng
Título da fonte: Repositório Institucional da UFSC
Texto Completo: https://repositorio.ufsc.br/handle/123456789/223064
Resumo: TCC(graduação) - Universidade Federal de Santa Catarina. Campus Araranguá. Engenharia da Computação.
id UFSC_150dc6031486df2a3dfe5b14f81b7bd1
oai_identifier_str oai:repositorio.ufsc.br:123456789/223064
network_acronym_str UFSC
network_name_str Repositório Institucional da UFSC
repository_id_str 2373
spelling Universidade Federal de Santa CatarinaHoeckler, PatrickOurique, Fabrício de Oliveira2021-05-17T14:06:38Z2021-05-17T14:06:38Z2021-05-05https://repositorio.ufsc.br/handle/123456789/223064TCC(graduação) - Universidade Federal de Santa Catarina. Campus Araranguá. Engenharia da Computação.Generative Adversarial Networks (GANs) are a subcategory of Artificial Neural Networks where the objective is the generation of new data, they do that by modeling the probability distribution of real data, usually coming from a dataset, and sampling from the modeled distribution in order to produce original data that is similar, and optimally indistinguish- able, from what was used in training. The principle behind GANs is based on a competition between two different networks, a discriminator who tries to distinguish real from fake data, and a generator who tries to fool the discriminator by producing data that is as close to the real one as possible. However, the competition between the networks makes training GANs be something notoriously difficult, instability and non-convergence are a common occurrence and many techniques have been proposed to improve not only the learning process, but also the quality of the generated results. The goal for this document was to analyse a number of the most common approaches and make an empirical evaluation of those, trying to apply the techniques in different datasets and seeing which configuration produces the best results. In the end there should be a roadmap that can be used to help guide the initial decisions about what method to use when constructing GANs for new and unknown situations.Generative Adversarial Networks (GANs) são uma subcategoria de Rede Neurais Artificiais onde o objetivo é a geração de novos dados, elas fazem isso tentando modelar a distribuição de probabilidades de dados reais, geralmente vindos de um dataset, e amostrando da distribuição modelada de modo a produzir dados originais que são similares, e idealmente indistinguíveis do que foi usado durante o treino. O princípio por trás de GANs é baseado em uma competição entre duas redes distintas, um discriminador que tenta distinguir entre dados reais e falsos, e um gerador que tenta enganar o discriminador produzindo dados que são o mais perto possível dos dados reais. Entretanto, a competição entre as duas redes faz do treinamento de GANs algo que é notoriamente difícil, instabilidade e não-convergência são ocorrências comuns e muitas técnicas foram propostas para melhorar não apenas o processo de aprendizado, mas também a qualidade dos resultados gerados. O objetivo deste documento foi de analisar um número de abordagens mais comuns e realizar uma avaliação empírica destas, tentando aplicar as técnicas em diferentes datasets e observando qual configuração produz os melhores resultados. Ao fim deve haver um roteiro que pode ser usado para ajudar a guiar as decisões iniciais sobre qual método utilizar ao construir GANs para novas situações desconhecidas.117Araranguá, SCEngenharia de ComputaçãoGenerative Adversarial NetworksGenerative ModelsDeep LearningNeural NetworksAn Analysis of Techniques for Building Generative Adversarial Networksinfo:eu-repo/semantics/publishedVersioninfo:eu-repo/semantics/bachelorThesisinfo:eu-repo/semantics/openAccessengreponame:Repositório Institucional da UFSCinstname:Universidade Federal de Santa Catarina (UFSC)instacron:UFSCLICENSElicense.txtlicense.txttext/plain; charset=utf-81383https://repositorio.ufsc.br/bitstream/123456789/223064/2/license.txt11ee89cd31d893362820eab7c4d46734MD52ORIGINALTCC.pdfTCC.pdftexto principalapplication/pdf34721159https://repositorio.ufsc.br/bitstream/123456789/223064/1/TCC.pdf98ae0c7e8e10bd91a720aa7b5634ae5fMD51123456789/2230642021-05-17 11:06:38.828oai:repositorio.ufsc.br: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ório de PublicaçõesPUBhttp://150.162.242.35/oai/requestopendoar:23732021-05-17T14:06:38Repositório Institucional da UFSC - Universidade Federal de Santa Catarina (UFSC)false
dc.title.pt_BR.fl_str_mv An Analysis of Techniques for Building Generative Adversarial Networks
title An Analysis of Techniques for Building Generative Adversarial Networks
spellingShingle An Analysis of Techniques for Building Generative Adversarial Networks
Hoeckler, Patrick
Engenharia de Computação
Generative Adversarial Networks
Generative Models
Deep Learning
Neural Networks
title_short An Analysis of Techniques for Building Generative Adversarial Networks
title_full An Analysis of Techniques for Building Generative Adversarial Networks
title_fullStr An Analysis of Techniques for Building Generative Adversarial Networks
title_full_unstemmed An Analysis of Techniques for Building Generative Adversarial Networks
title_sort An Analysis of Techniques for Building Generative Adversarial Networks
author Hoeckler, Patrick
author_facet Hoeckler, Patrick
author_role author
dc.contributor.pt_BR.fl_str_mv Universidade Federal de Santa Catarina
dc.contributor.author.fl_str_mv Hoeckler, Patrick
dc.contributor.advisor1.fl_str_mv Ourique, Fabrício de Oliveira
contributor_str_mv Ourique, Fabrício de Oliveira
dc.subject.por.fl_str_mv Engenharia de Computação
Generative Adversarial Networks
Generative Models
Deep Learning
Neural Networks
topic Engenharia de Computação
Generative Adversarial Networks
Generative Models
Deep Learning
Neural Networks
description TCC(graduação) - Universidade Federal de Santa Catarina. Campus Araranguá. Engenharia da Computação.
publishDate 2021
dc.date.accessioned.fl_str_mv 2021-05-17T14:06:38Z
dc.date.available.fl_str_mv 2021-05-17T14:06:38Z
dc.date.issued.fl_str_mv 2021-05-05
dc.type.status.fl_str_mv info:eu-repo/semantics/publishedVersion
dc.type.driver.fl_str_mv info:eu-repo/semantics/bachelorThesis
format bachelorThesis
status_str publishedVersion
dc.identifier.uri.fl_str_mv https://repositorio.ufsc.br/handle/123456789/223064
url https://repositorio.ufsc.br/handle/123456789/223064
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.format.none.fl_str_mv 117
dc.publisher.none.fl_str_mv Araranguá, SC
publisher.none.fl_str_mv Araranguá, SC
dc.source.none.fl_str_mv reponame:Repositório Institucional da UFSC
instname:Universidade Federal de Santa Catarina (UFSC)
instacron:UFSC
instname_str Universidade Federal de Santa Catarina (UFSC)
instacron_str UFSC
institution UFSC
reponame_str Repositório Institucional da UFSC
collection Repositório Institucional da UFSC
bitstream.url.fl_str_mv https://repositorio.ufsc.br/bitstream/123456789/223064/2/license.txt
https://repositorio.ufsc.br/bitstream/123456789/223064/1/TCC.pdf
bitstream.checksum.fl_str_mv 11ee89cd31d893362820eab7c4d46734
98ae0c7e8e10bd91a720aa7b5634ae5f
bitstream.checksumAlgorithm.fl_str_mv MD5
MD5
repository.name.fl_str_mv Repositório Institucional da UFSC - Universidade Federal de Santa Catarina (UFSC)
repository.mail.fl_str_mv
_version_ 1766805391996354560

Registros relacionados