Transcription of Wasserstein Generative Adversarial Networks
{{id}} {{{paragraph}}}
Example: barber
Wasserstein Generative Adversarial Networks
Wasserstein Generative Adversarial NetworksMartin Arjovsky1 Soumith Chintala2L eon Bottou1 2 AbstractWe introduce a new algorithm named WGAN,an alternative to traditional GAN training. Inthis new model, we show that we can improvethe stability of learning, get rid of problems likemode collapse, and provide meaningful learningcurves useful for debugging and hyperparametersearches. Furthermore, we show that the cor-responding optimization problem is sound, andprovide extensive theoretical work highlightingthe deep connections to different distances be-tween IntroductionThe problem this paper is concerned with is that of unsu-pervised learning. Mainly, what does it mean to learn aprobability distribution? The classical answer to this is tolearn a probability density. This is often done by defininga parametric family of densities(P ) Rdand finding theone that maximized the likelihood on our data: if we havereal data examples{x(i)}mi=1, we would solve the problemmax Rd1mm i=1logP (x(i))If the real data distributionPradmits a density andP is thedistribution of the parametrized densityP , then, asymp-totically, this amounts to minimizing the Kullback-LeiblerdivergenceKL(Pr P ).
Wasserstein Generative Adversarial Networks Figure 1: These plots show ˆ(P ;P 0) as a function of when ˆis the EM distance (left plot) or the JS divergence (right plot).The EM plot is continuous and provides a usable gradient everywhere.
Loading..
Tags:
Information
Domain:
Source:
Link to this page:
Please notify us if you found a problem with this document: