Timezone: »

Banach Wasserstein GAN
Jonas Adler · Sebastian Lunz

Tue Dec 04 07:45 AM -- 09:45 AM (PST) @ Room 210 #18
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.

Author Information

Jonas Adler (KTH - Royal Institute of Technology)

I’m a Research Scientist at Elekta, pursuing a PhD in Applied Mathematics working under the supervision of Ozan Öktem. I do research in inverse problems and machine learning, especially focusing on the intersection between model-driven and data-driven methods. Organizing [DLIP2019](https://sites.google.com/view/dlip2019).

Sebastian Lunz (University of Cambridge)

More from the Same Authors