How (not) to Train your Generative Model: Scheduled Sampling, Likelihood, Adversary?

Ferenc Huszár

Modern applications and progress in deep learning research have created renewed interest for generative models of text and of images. However, even today it is unclear what objecti…