Generative adversarial networks (GANs) provide a way to learn deeprepresentations without extensively annotated training data. They achieve thisthrough deriving backpropagation signals through a competitive processinvolving a pair of networks. The representations that can be learned by GANsmay be used in a variety of applications, including image synthesis, semanticimage editing, style transfer, image super-resolution and classification. Theaim of this review paper is to provide an overview of GANs for the signalprocessing community, drawing on familiar analogies and concepts wherepossible. In addition to identifying different methods for training andconstructing GANs, we also point to remaining challenges in their theory andapplication.
translated by 谷歌翻译