生成式对抗网络研究综述

Generative adversarial network: An overview

深度学习领域一个十分活跃的分支—生成式对抗网络(GAN)已经成为人工智能学界一个热门的研究方向。生成式对抗网络采用无监督的学习方式,自动从源数据中进行学习,在不需要人工对数据集进行标记的情况下就可以产生令人惊叹的效果。阐述了GAN的背景、基本思想,对其相关理论、训练机制和应用研究进行了梳理,总结了GAN的常见网络构架、训练技巧与模型评估标准,还进行了GAN与其他生成模型VAE、衍生模型的对比,最后进行分析总结,指出GAN的优缺点并对未来发展方向进行展望。

Generative adversarial network(GAN)is an active branch of deep learning field,which has become a popular research direction in the field of artificial intelligence.GAN adopts an unsupervised learning method and automatically learns from the source data,which can produce amazing effects without artificially labeling data.In this paper,we present the background,basic idea of GAN and comb its related theory,training mechanism and state-of-the-art applications.We also summarize the common network architectures,training skills and model evaluation metrics,and compareGAN with other generative model VAE and GAN variants.Finally,we point out the advantages and disadvantages of the GAN and look forward to the further development direction.