Current browse context:
cs
Change to browse by:
References & Citations
Computer Science > Computer Vision and Pattern Recognition
Title: Banach Wasserstein GAN
(Submitted on 18 Jun 2018 (v1), last revised 11 Jan 2019 (this version, v2))
Abstract: Wasserstein Generative Adversarial Networks (WGANs) can be used to generate realistic samples from complicated image distributions. The Wasserstein metric used in WGANs is based on a notion of distance between individual images, which induces a notion of distance between probability distributions of images. So far the community has considered $\ell^2$ as the underlying distance. We generalize the theory of WGAN with gradient penalty to Banach spaces, allowing practitioners to select the features to emphasize in the generator. We further discuss the effect of some particular choices of underlying norms, focusing on Sobolev norms. Finally, we demonstrate a boost in performance for an appropriate choice of norm on CIFAR-10 and CelebA.
Submission history
From: Jonas Adler [view email][v1] Mon, 18 Jun 2018 12:15:43 GMT (5556kb,D)
[v2] Fri, 11 Jan 2019 12:57:46 GMT (9183kb,D)
Link back to: arXiv, form interface, contact.