Workshop on Machine Learning in Speech and Language Processing

September 13, 2016
San Francisco, CA, USA
Speaker: Ian Goodfellow (OpenAI)

Title: Generative Adversarial Networks

Abstract:
Generative adversarial networks (GANs) are a recently introduced class of generative models. They do not provide density estimates, but they do provide high-quality samples. Conditional GANs can also produce high-quality samples given some input. Because of these properties, GANs hold promise for text-to-speech. Like many other deep generative models, GANs are universal approximators of probability distributions. Unlike many other generative models, they do not require the estimation of an intractable normalization constant, the use of Markov chain approximations, or variational approximations. Instead, they use supervised learning techniques to estimate the ratio between the model density and the data density, and use this ratio to estimate the gradient on the model parameters.