First Order Generative Adversarial Networks

ICML 2018 Calvin SewardThomas UnterthinerUrs BergmannNikolay JetchevSepp Hochreiter

GANs excel at learning high dimensional distributions, but they can update generator parameters in directions that do not correspond to the steepest descent direction of the objective. Prominent examples of problematic update directions include those used in both Goodfellow's original GAN and the WGAN-GP... (read more)

PDF Abstract
TASK DATASET MODEL METRIC NAME METRIC VALUE GLOBAL RANK RESULT BENCHMARK
Image Generation CIFAR-10 FOGAN FID 27.4 # 22
Image Generation LSUN Bedroom 256 x 256 FOGAN FID 11.4 # 8

Methods used in the Paper