Deep generative models: Survey

Abstract

Generative models have found their way to the forefront of deep learning the last decade and so far, it seems that the hype will not fade away any time soon. In this paper, we give an overview of the most important building blocks of most recent revolutionary deep generative models such as RBM, DBM, DBN, VAE and GAN. We will also take a look at three of state-of-the-art generative models, namely PixelRNN, DRAW and NADE. We will delve into their unique architectures, the learning procedures and their potential and limitations. We will also review some of the known issues that arise when trying to design and train deep generative architectures using shallow ones and how different models deal with these issues. This paper is not meant to be a comprehensive study of these models, but rather a starting point for those who bear an interest in the field.