{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/gans-beyond-divergence-minimization","title":"GANs beyond divergence minimization","arxiv_id":"1809.02145","date":"2018-09-06","proceeding":null,"authors":["Alexia Jolicoeur-Martineau"],"abstract":"Generative adversarial networks (GANs) can be interpreted as an adversarial\ngame between two players, a discriminator D and a generator G, in which D\nlearns to classify real from fake data and G learns to generate realistic data\nby \"fooling\" D into thinking that fake data is actually real data. Currently, a\ndominating view is that G actually learns by minimizing a divergence given that\nthe general objective function is a divergence when D is optimal. However, this\nview has been challenged due to inconsistencies between theory and practice. In\nthis paper, we discuss of the properties associated with most loss functions\nfor G (e.g., saturating/non-saturating f-GAN, LSGAN, WGAN, etc.). We show that\nthese loss functions are not divergences and do not have the same equilibrium\nas expected of divergences. This suggests that G does not need to minimize the\nsame objective function as D maximize, nor maximize the objective of D after\nswapping real data with fake data (non-saturating GAN) but can instead use a\nwide range of possible loss functions to learn to generate realistic data. We\ndefine GANs through two separate and independent D maximization and G\nminimization steps. We generalize the generator step to four new classes of\nloss functions, most of which are actual divergences (while traditional G loss\nfunctions are not). We test a wide variety of loss functions from these four\nclasses on a synthetic dataset and on CIFAR-10. We observe that most loss\nfunctions converge well and provide comparable data generation quality to\nnon-saturating GAN, LSGAN, and WGAN-GP generator loss functions, whether we use\ndivergences or non-divergences. These results suggest that GANs do not conform\nwell to the divergence minimization theory and form a much broader range of\nmodels than previously assumed.","url_abs":"http://arxiv.org/abs/1809.02145v1","url_pdf":"http://arxiv.org/pdf/1809.02145v1.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"gans-beyond-divergence-minimization","repo_url":"https://github.com/AlexiaJM/GANsBeyondDivergenceMin","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok"}}],"tasks":[],"methods":[{"method_slug":"batch-normalization","method_name":"Batch Normalization"},{"method_slug":"convolution","method_name":"Convolution"},{"method_slug":"dense-connections","method_name":"Dense Connections"},{"method_slug":"gan-least-squares-loss","method_name":"GAN Least Squares Loss"},{"method_slug":"lsgan","method_name":"LSGAN"},{"method_slug":"relu","method_name":"ReLU"},{"method_slug":"wgan","method_name":"WGAN"}],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":null,"mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}