Learning disconnected manifolds: no GAN's land - Inria - Institut national de recherche en sciences et technologies du numérique Accéder directement au contenu
Communication Dans Un Congrès Année : 2020

Learning disconnected manifolds: no GAN's land

Ugo Tanielian
  • Fonction : Auteur
  • PersonId : 1029602
Thibaut Issenhuth
  • Fonction : Auteur
Elvis Dohmatob
  • Fonction : Auteur
Jérémie Mary

Résumé

Typical architectures of Generative Adversarial Networks make use of a unimodal latent/input distribution transformed by a continuous generator. Consequently, the modeled distribution always has connected support which is cumbersome when learning a disconnected set of manifolds. We formalize this problem by establishing a "no free lunch" theorem for the disconnected manifold learning stating an upper-bound on the precision of the targeted distribution. This is done by building on the necessary existence of a low-quality region where the generator continuously samples data between two disconnected modes. Finally, we derive a rejection sampling method based on the norm of generator's Jacobian and show its efficiency on several generators including BigGAN.
Fichier principal
Vignette du fichier
main.pdf (6.22 Mo) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-02778441 , version 1 (05-06-2020)

Identifiants

Citer

Ugo Tanielian, Thibaut Issenhuth, Elvis Dohmatob, Jérémie Mary. Learning disconnected manifolds: no GAN's land. International Conference on Machine Learning 2020, Jul 2020, Vienna, Austria. ⟨hal-02778441⟩
29 Consultations
17 Téléchargements

Altmetric

Partager

Gmail Facebook X LinkedIn More