R\'{e}nyi Generative Adversarial Networks

Unsupervised learning is the problem of educing information from a large unlabeled dataset and, in the context of generative models, is a relatively new area that has received much attention. Two prominent objectives in generative modeling consist of determining the underlying probability distribution function of a dataset or generating data that mimics it. Classical techniques for the former include maximum likelihood estimators, methods of moments estimators and Bayesian estimators. The main approaches for the latter include generative adversarial networks (GANs) [15], [5], [36], [10], autoencoders/variational autoencoders (VAEs) [22], generative autoregressive models [34], invertible flow based latent vector models [23], or hybrids of the above models [16]. Compared to other approaches, GANs have garnered the most interest (e.g., see surveys in [10], [43], [44]); unlike density estimation models, GANs can efficiently represent distributions confined to a low dimensional manifold [5] and are therefore the focus of this paper. Prior Work: The original GAN [15] consists of a generative neural network competing with a discriminatory neural network in a min-max game. GANs were enhanced with the introduction of deep convolutional GANs (DCGANs) [36] which use convolutional layers to learn higher dimensional dependencies that are inherent in complex datasets such as images [36].