Unveiling the Latent Space Geometry of Push-Forward Generative Models

Thibaut Issenhuth, Ugo Tanielian, Jérémie Mary, David Picard

Abstract: Many deep generative models are defined as a push-forward of a Gaussian measure by a continuous generator, such as Generative Adversarial Networks (GANs) or Variational Auto-Encoders (VAEs). This w...