VQ-VAE

E755720

neural network model

VQ-VAE is a neural network model that combines vector quantization with variational autoencoders to learn discrete latent representations for tasks like image and audio generation.

Try in SPARQL Jump to: Surface forms Statements Referenced by

Observed surface forms (1)

Surface form	Occurrences
VQ-VAE-2	1

Statements (47)

Predicate	Object
instanceOf	neural network model ⓘ
addressesProblem	learning discrete representations ⓘ posterior collapse in VAEs ⓘ
basedOn	variational autoencoder ⓘ
canBeExtendedTo	VQ-VAE-2 NERFINISHED ⓘ hierarchical VQ-VAE NERFINISHED ⓘ
codebookSize	hyperparameter ⓘ
embeddingDimension	hyperparameter ⓘ
fullName	Vector Quantized Variational Autoencoder NERFINISHED ⓘ
hasAdvantage	avoids sampling from continuous latent distributions at training time ⓘ enables use of powerful autoregressive priors over codes ⓘ produces interpretable discrete codes ⓘ
hasComponent	codebook ⓘ codebook loss term ⓘ commitment loss term ⓘ decoder ⓘ embedding vectors ⓘ encoder ⓘ reconstruction loss term ⓘ
hasLatentSpaceType	discrete latent space ⓘ
inputType	audio waveforms ⓘ images ⓘ spectrograms ⓘ
inspired	subsequent discrete representation models ⓘ
introducedInPaper	Neural Discrete Representation Learning NERFINISHED ⓘ
latentRepresentation	indices into a codebook of embeddings ⓘ
outputType	reconstructed audio ⓘ reconstructed images ⓘ
primaryApplication	audio generation ⓘ compression ⓘ image generation ⓘ representation learning ⓘ speech generation ⓘ
proposedBy	Aaron van den Oord NERFINISHED ⓘ Koray Kavukcuoglu NERFINISHED ⓘ Oriol Vinyals NERFINISHED ⓘ
publicationYear	2017 ⓘ
publishedByOrganization	DeepMind NERFINISHED ⓘ
usedWith	PixelCNN prior NERFINISHED ⓘ WaveNet prior NERFINISHED ⓘ
usesOptimizationMethod	Adam optimizer NERFINISHED ⓘ stochastic gradient descent ⓘ
usesTechnique	vector quantization ⓘ
usesTrainingObjective	codebook vector quantization ⓘ commitment loss regularization ⓘ reconstruction error minimization ⓘ
usesTrick	straight-through estimator ⓘ

Referenced by (2)

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

Aaron van den Oord → developed → VQ-VAE ⓘ

this entity surface form: VQ-VAE-2