{"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/many-paths-to-equilibrium-gans-do-not-need-to","title":"Many Paths to Equilibrium: GANs Do Not Need to Decrease a Divergence At Every Step","arxiv_id":"1710.08446","date":"2017-10-23","proceeding":"ICLR 2018 1","authors":["William Fedus","Mihaela Rosca","Balaji Lakshminarayanan","Andrew M. Dai","Shakir Mohamed","Ian Goodfellow"],"abstract":"Generative adversarial networks (GANs) are a family of generative models that\ndo not minimize a single training criterion. Unlike other generative models,\nthe data distribution is learned via a game between a generator (the generative\nmodel) and a discriminator (a teacher providing training signal) that each\nminimize their own cost. GANs are designed to reach a Nash equilibrium at which\neach player cannot reduce their cost without changing the other players'\nparameters. One useful approach for the theory of GANs is to show that a\ndivergence between the training distribution and the model distribution obtains\nits minimum value at equilibrium. Several recent research directions have been\nmotivated by the idea that this divergence is the primary guide for the\nlearning process and that every step of learning should decrease the\ndivergence. We show that this view is overly restrictive. During GAN training,\nthe discriminator provides learning signal in situations where the gradients of\nthe divergences between distributions would not be useful. We provide empirical\ncounterexamples to the view of GAN training as divergence minimization.\nSpecifically, we demonstrate that GANs are able to learn distributions in\nsituations where the divergence minimization point of view predicts they would\nfail. We also show that gradient penalties motivated from the divergence\nminimization perspective are equally helpful when applied in other contexts in\nwhich the divergence minimization perspective does not predict they would be\nhelpful. This contributes to a growing body of evidence that GAN training may\nbe more usefully viewed as approaching Nash equilibria via trajectories that do\nnot necessarily minimize a specific divergence at each step.","url_abs":"http://arxiv.org/abs/1710.08446v3","url_pdf":"http://arxiv.org/pdf/1710.08446v3.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":"many-paths-to-equilibrium-gans-do-not-need-to","repo_url":"https://github.com/kodalinaveen3/DRAGAN","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok","spdx":"MIT"}}],"tasks":[],"methods":[{"method_slug":"convolution","method_name":"Convolution"}],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1710.08446","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}