Failure Modes of Variational Autoencoders and Their Effects on Downstream Tasks

Variational Auto-encoders (VAEs) are deep generative latent variable models that are widely used for a number of downstream tasks. While it has been demonstrated that VAE training can suffer from a number of pathologies, existing literature lacks characterizations of exactly when these pathologies o...

Descripción completa

Detalles Bibliográficos

Autores principales:	Yacoby, Yaniv, Pan, Weiwei, Doshi-Velez, Finale
Formato:	Artículo
Publicado:	2020
Materias:	Machine Learning
Acceso en línea:	https://arxiv.org/abs/2007.07124 http://arxiv.org/abs/2007.07124
Etiquetas:	Agregar Etiqueta Sin Etiquetas, Sea el primero en etiquetar este registro!