Wasserstein GAN

E290870

deep generative model generative adversarial network variant machine learning model neural network architecture

Wasserstein GAN is a variant of generative adversarial networks that improves training stability and sample quality by optimizing the Wasserstein (Earth Mover’s) distance between real and generated data distributions.

Try in SPARQL Jump to: Surface forms Statements Referenced by

All labels observed (3)

Label	Occurrences
WGAN	2
Wasserstein GAN canonical	2
WGAN-GP	1

Statements (47)

Predicate	Object
instanceOf	deep generative model ⓘ generative adversarial network variant ⓘ machine learning model ⓘ neural network architecture ⓘ
advantage	correlates generator loss with sample quality ⓘ provides smoother loss landscape ⓘ reduces mode collapse compared to vanilla GANs ⓘ
alsoKnownAs	Wasserstein GAN ⓘ surface form: WGAN
basedOn	Earth Mover's distance ⓘ Wasserstein distance ⓘ
category	adversarial learning method ⓘ unsupervised learning method ⓘ
comparedTo	original GAN ⓘ
contrastedWith	Jensen–Shannon divergence in original GAN ⓘ
criticObjective	maximize difference between scores on real and fake samples ⓘ
criticRole	estimates Wasserstein distance ⓘ
criticUpdateCount	multiple critic steps per generator step ⓘ
distanceType	Wasserstein-1 metric ⓘ
domainOfApplication	audio generation ⓘ image generation ⓘ representation learning ⓘ text generation ⓘ
enforcesLipschitzConstraintBy	weight clipping ⓘ
evaluationProperty	loss remains informative during training ⓘ
generatorObjective	minimize critic score on generated samples ⓘ
implementationDetail	often uses RMSProp or Adam for optimization ⓘ often uses weight clipping to a small range like [-0.01, 0.01] ⓘ
inspiredFollowUpModel	Wasserstein GAN self-linksurface differs ⓘ surface form: WGAN-GP improved WGAN with gradient penalty ⓘ
mathematicalFoundation	optimal transport theory ⓘ
optimizes	Wasserstein-1 distance between real and generated distributions ⓘ
primaryGoal	improve GAN training stability ⓘ improve sample quality ⓘ provide meaningful loss metric for GANs ⓘ
proposedBy	Léon Bottou ⓘ Martin Arjovsky ⓘ Soumith Chintala ⓘ
proposedInPaper	Wasserstein GAN self-link ⓘ
publicationYear	2017 ⓘ
relatedConcept	Lipschitz continuity ⓘ gradient penalty ⓘ mode collapse in GANs ⓘ
replacesComponent	discriminator with critic ⓘ
trainingProcedure	alternates critic and generator updates ⓘ
trainingProperty	critic is constrained to be 1-Lipschitz ⓘ critic outputs real-valued scores instead of probabilities ⓘ
usesLossFunction	Wasserstein loss ⓘ

Referenced by (5)

Full triples — surface form annotated when it differs from this entity's canonical label.

Generative Adversarial Networks → notableVariant → Wasserstein GAN ⓘ

this entity surface form: WGAN

Wasserstein GAN → alsoKnownAs → Wasserstein GAN ⓘ

this entity surface form: WGAN

Wasserstein GAN → proposedInPaper → Wasserstein GAN self-link ⓘ

Wasserstein GAN → inspiredFollowUpModel → Wasserstein GAN self-linksurface differs ⓘ

this entity surface form: WGAN-GP