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

Generative models, especially Generative Adversarial Networks (GANs), havereceived significant attention recently. However, it has been observed that interms of some attributes, e.g. the number of simple geometric primitives in animage, GANs are not able to learn the target distribution in practice.Motivated by this observation, we discover two specific problems of GANsleading to anomalous generalization behaviour, which we refer to as the sampleinsufficiency and the pixel-wise combination. For the first problem of sampleinsufficiency, we show theoretically and empirically that the batchsize of thetraining samples in practice may be insufficient for the discriminator to learnan accurate discrimination function. It could result in unstable trainingdynamics for the generator, leading to anomalous generalization. For the secondproblem of pixel-wise combination, we find that besides recognizing thepositive training samples as real, under certain circumstances, thediscriminator could be fooled to recognize the pixel-wise combinations (e.g.pixel-wise average) of the positive training samples as real. However, thosecombinations could be visually different from the real samples in the targetdistribution. With the fooled discriminator as reference, the generator wouldobtain biased supervision further, leading to the anomalous generalizationbehaviour. Additionally, in this paper, we propose methods to mitigate theanomalous generalization of GANs. Extensive experiments on benchmark show ourproposed methods improve the FID score up to 30\% on natural image dataset.

On the Anomalous Generalization of GANs