Generative Adversarial Networks: An Overview

This overview explains generative adversarial networks, which provide a way to learn deep representations without extensively annotated training data by pitting a pair of networks against each other in a competitive process that produces backpropagation signals. Written for the signal processing community, the article draws on familiar analogies and concepts to make GANs accessible, and it identifies the different methods available for training and constructing these models.

The paper surveys the range of applications enabled by GAN-learned representations, including image synthesis, semantic image editing, style transfer, image super-resolution, and classification. It also points to the challenges that remain in both the theory and application of GANs, making it a useful entry point for practitioners seeking to understand the promise and limitations of adversarial generative modeling.