Flexible and accurate inference and learning for deep generative models

Helmholtz machine and later variational autoencoder algorithms (but unlike adversarial methods) our approach learns an explicit inference or "recognition" model