生成对抗网络在各领域应用研究进展

随着深度学习的快速发展, 生成式模型领域也取得了显著进展. 生成对抗网络(Generative adversarial network, GAN)是一种无监督的学习方法, 它是根据博弈论中的二人零和博弈理论提出的. GAN具有一个生成器网络和一个判别器网络, 并通过对抗学习进行训练. 近年来, GAN成为一个炙手可热的研究方向. GAN不仅在图像领域取得了不错的成绩, 还在自然语言处理(Natural language processing, NLP)以及其他领域崭露头角. 本文对GAN的基本原理、训练过程和传统GAN存在的问题进行了阐述, 进一步详细介绍了通过损失函数的修改、网络结构的变化以及两者结合的手段提出的GAN变种模型的原理结构, 其中包括: 条件生成对抗网络(Conditional GAN, CGAN)、基于Wasserstein 距离的生成对抗网络(Wasserstein-GAN, WGAN)及其基于梯度策略的WGAN (WGAN-gradient penalty, WGAN-GP)、基于互信息理论的生成对抗网络(Informational-GAN, InfoGAN)、序列生成对抗网络(Sequence GAN, SeqGAN)、Pix2Pix、循环一致生成对抗网络(Cycle-consistent GAN, Cycle GAN)及其增强Cycle-GAN (Augmented CycleGAN). 概述了在计算机视觉、语音与NLP领域中基于GAN和相应GAN变种模型的基本原理结构, 其中包括: 基于CGAN的脸部老化应用(Face aging CGAN, Age-cGAN)、双路径生成对抗网络(Two-pathway GAN, TP-GAN)、表示解析学习生成对抗网络(Disentangled representation learning GAN, DR-GAN)、对偶学习生成对抗网络(DualGAN)、GeneGAN、语音增强生成对抗网络(Speech enhancement GAN, SEGAN)等. 介绍了GAN在医学、数据增强等领域的应用情况, 其中包括: 数据增强生成对抗网络(Data augmentation GAN, DAGAN)、医学生成对抗网络(Medical GAN, MedGAN)、无监督像素级域自适应方法(Unsupervised pixel-level domain adaptation method, PixelDA). 最后对GAN未来发展趋势及方向进行了展望.

Research Progress on Application of Generative Adversarial Networks in Various Fields

With the rapid development of deep learning, the field of generative models has also made significant progress. Generative adversarial network (GAN) is an unsupervised learning method based on the zero-sum game theory in game theory. GAN has a generator network and a discriminator network and trains through adversarial learning. In the past two years, GAN has become a hot research direction. GAN has not only achieved good results in the field of computer vision, but also emerged in natural language processing (NLP) and other fields. This paper expounds the basic principles of GAN, the training process and the problems existing in traditional GAN, and further introduces the principal structure of the GAN variant model proposed by the modification of the loss function, the change of the network structure and the combination of the two, e.g., conditional GAN (CGAN), Wasserstein-GAN (WGAN), WGAN-gradient penalty (WGAN-GP), informational GAN (InfoGAN), sequence GAN (SeqGAN), Pix2Pix, cycle-consistent GAN (CycleGAN) and augmented CycleGAN, and so on. Then in the areas of computer vision, speech synthetics and analysis and NLP, we review the structure of the principle networks and models, including Age-cGAN for face aging, two-pathway GAN (TP-GAN), disentangled representation learning GAN (DR-GAN), DualGAN, GeneGAN, speech enhancement GAN (SEGAN), gumbel-softmax GAN, and so forth. Then we also introduce the applications of GAN in the field of medicine, data enhancement,etc, including data augmentation GAN (DAGAN), medical GAN (MedGAN), unsupervised pixel-level domain adaptation method (PixelDA), and so on. Finally, the future trends and directions of GAN are prospected.