English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Modern Generative Adversarial Networks (GANs) generate realistic imagesremarkably well. Previous work has demonstrated the feasibility of"GAN-classifiers" that are distinct from the co-trained discriminator, andoperate on images generated from a frozen GAN. That such classifiers work atall affirms the existence of "knowledge gaps" (out-of-distribution artifactsacross samples) present in GAN training. We iteratively train GAN-classifiersand train GANs that "fool" the classifiers (in an attempt to fill the knowledgegaps), and examine the effect on GAN training dynamics, output quality, andGAN-classifier generalization. We investigate two settings, a small DCGANarchitecture trained on low dimensional images (MNIST), and StyleGAN2, a SOTAGAN architecture trained on high dimensional images (FFHQ). We find that theDCGAN is unable to effectively fool a held-out GAN-classifier withoutcompromising the output quality. However, StyleGAN2 can fool held-outclassifiers with no change in output quality, and this effect persists overmultiple rounds of GAN/classifier training which appears to reveal an orderingover optima in the generator parameter space. Finally, we study differentclassifier architectures and show that the architecture of the GAN-classifierhas a strong influence on the set of its learned artifacts.

Sequential training of GANs against GAN-classifiers reveals correlated "knowledge gaps" present among independently trained GAN instances