litlas
Log inCreate account

Generalization and Equilibrium in Generative Adversarial Nets (GANs)

Authors: Sanjeev Arora, Rong Ge, Yingyu Liang, Tengyu Ma, Yi Zhang - International Conference on Machine Learning, ICML 2017 cited by 726

We show that training of generative adversarial network (GAN) may not have good generalization properties; e.g., training may appear successful but the trained distribution may be far from target distribution in standard metrics. However, generalization does occur for a weaker metric called neural net distance. It is also shown that an approximate pure equilibrium exists in the discriminator/generator game for a special class of generators with natural training objectives when generator capacity and training set sizes are moderate. This existence of equilibrium inspires MIX+GAN protocol, which can be combined with any existing GAN training, and empirically shown to improve some of them.

NIPS 2016 Tutorial:Generative Adversarial…NIPS 2016 Tutorial: Generative Adversarial NetworksLearning to ProtectCommunications with…Learning to Protect Communications with Adversarial Neural CryptographyAdaGAN: BoostingGenerative ModelsAdaGAN: Boosting Generative ModelsDo GANs actually learnthe distribution? An…Do GANs actually learn the distribution? An empirical studyConditional ImageSynthesis With Auxiliar…Conditional Image Synthesis With Auxiliary Classifier GANsWasserstein GANWasserstein GANBEGAN: BoundaryEquilibrium Generative…BEGAN: Boundary Equilibrium Generative Adversarial NetworksOn theDiscrimination-Generali…On the Discrimination-Generalization Tradeoff in GANsDemystifying MMD GANsDemystifying MMD GANsApproximability ofDiscriminators Implies…Approximability of Discriminators Implies Diversity in GANsGANs May Have No NashEquilibriaGANs May Have No Nash EquilibriaOn How Well GenerativeAdversarial Networks…On How Well Generative Adversarial Networks Learn Densities: Nonparametric and Parametric ResultsMinmax Optimization:Stable Limit Points of…Minmax Optimization: Stable Limit Points of Gradient Descent Ascent are Locally OptimalDeconstructingGenerative Adversarial…Deconstructing Generative Adversarial NetworksGenerative AdversarialNetworks (GANs)…Generative Adversarial Networks (GANs): Challenges, Solutions, and Future DirectionsA UniversalApproximation Theorem o…A Universal Approximation Theorem of Deep Neural Networks for Expressing Probability DistributionsTrain simultaneously,generalize better…Train simultaneously, generalize better: Stability of gradient-based minimax learnersGAT-GMM: GenerativeAdversarial Training fo…GAT-GMM: Generative Adversarial Training for Gaussian Mixture ModelsGeneralization andEquilibrium in…Generalization and Equilibrium in Generative Adversarial Nets (GANs)Earlier referencesFocus paperCiting papersOlderNewer

Click a node to pin it, click the empty canvas to go back to this paper, or hover to preview. Open a node’s page from its title.