A Quantitative Measure of Generative Adversarial Network Distributions

Dan Hendrycks*, Steven Basart*

Feb 17, 2017 (modified: Mar 20, 2017) ICLR 2017 workshop submission readers: everyone
  • Abstract: We introduce a new measure for evaluating the quality of distributions learned by Generative Adversarial Networks (GANs). This measure computes the Kullback-Leibler divergence from a GAN-generated image set to a real image set. Since our measure utilizes a GAN's whole distribution, our measure penalizes outputs lacking in diversity, and it contrasts with evaluating GANs based upon a few cherry-picked examples. We demonstrate the measure's efficacy on the MNIST, SVHN, and CIFAR-10 datasets.
  • Conflicts: uchicago.edu, ttic.edu