Faculty Recruiting Support CICS

Better GANs by Using Kernels

01 Oct
Tuesday, 10/01/2019 10:00am to 11:00am
Computer Science Building, Room 151

Generative adversarial networks have led to huge improvements in sample quality for image generation. But their success is hindered by both practical and theoretical problems, leading to the proposal of a huge number of alternative methods over the last few years. We study one class of alternatives, the MMD GAN, which uses a similar architecture to an original GAN but does some of its optimization in closed form, in a Hilbert space. We deepen the understanding of these models, with a particular focus on the behavior of gradient penalties in this context. Based on this, we propose a method to constrain the gradient analytically, rather than with an additive optimization penalty. Our new method, the Scaled MMD GAN, enjoys pleasing theoretical properties and also achieves excellent unsupervised image generation on CelebA and ImageNet.

Based on joint work with Michael Arbel, Mikolaj Binkowski, and Arthur Gretton.

Dougal Sutherland is a Research Assistant Professor at TTIC, and will begin as an Assistant Professor in UBC Computer Science in 2020. Dougal received a PhD from CMU in 2016 and was a postdoc at the Gatsby Unit, UCL from 2016-19. Dougal's research focuses on measuring and understanding differences between distributions, with applications including two-sample testing, generative models, and distribution regression. These areas, in addition to being of independent interest, provide a nice testbed for nontrivial combinations of the advantages of kernel methods with those of deep learning.


Faculty Host