https://www.toutiao.com/a6708994742929064451/
Formula confrontation Network (Generative adversarial networks, GAN) is one of the hotspots of the current academic artificial intelligence the most important. The main idea is to set a zero-sum game, two players to achieve learning through confrontation. A player in the game is called for the generator, its main job is to generate a sample, and try to make it appear to be consistent with the training samples. in addition, a player known as a discriminator, its purpose is to accurately determine whether the input sample real training samples. a common analogy is to imagine these two networks to counterfeiters and police. GAN training process is similar to counterfeiters counterfeit production level as much as possible to improve the fool police, who are constantly improving the ability to identify to identify counterfeit as of GAN continued training, the ability to counterfeit producers and the police will continue to improve.
Formula confrontation Network
Compared to the previous generation model, GAN model has the following obvious advantages: First, the data generated by the linear dimension of complexity associated with, for large sample dimensions generated only increase the output dimensions neural network, unlike traditional We face the same calculation model rose; Second data distribution limits is not dominant, thereby eliminating the need for manual design model distribution; third generation GAN handwritten numbers, face, CIFAR-10 than the other samples VAE , PixelCNN and other generation model clearer.
GAN outstanding ability to generate not only be used to generate various types of image and natural language data, but also inspired and promoted the development of various types of semi-supervised learning and unsupervised learning task. GAN combination, the researchers fill in data, image translation, data synthesis, imitation learning, and many have made a breakthrough.
GAN conventional methods and data padding effect [3]
iGAN generating sample [4]
FIG FIG Translation [5]
Manipulator using synthetic data training GAN [6]
However, the original GAN model also has a number of problems, including the difficulties of convergence, can not generate discrete data, it is difficult evaluation. In this paper, the recent development of GAN reviewed for GAN improvements in both generating mechanism, identification mechanisms were introduced, and sorts out its field of application. discusses the relationship between GAN and parallel thinking on this basis.
Organizational Structure
references
[1] I. Goodfellow et al., “Generative adversarial nets,” in Advances in neural information processing systems, 2014, pp. 2672–2680
[2] I. Goodfellow, “NIPS 2016 Tutorial: Generative Adversarial Networks,” arXiv:1701.00160 [cs], Dec. 2016(arXiv: 1701.00160)
[3] S. Iizuka, E. Simo-Serra, and H. Ishikawa, “Globally and locally consistent image completion,” ACM Transactions on Graphics, vol. 36, no. 4, pp. 1–14, Jul. 2017
[4] J.-Y. Zhu, P. Krähenbühl, E. Shechtman, and A. A. Efros, “Generative Visual Manipulation on the Natural Image Manifold,” in European Conference on Computer Vision, 2016, pp. 597–613
[5] P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros, “Image-to-image translation with conditional adversarial networks,” arXiv preprint arXiv:1611.07004, 2016
[6] K. Bousmalis et al., “Using Simulation and Domain Adaptation to Improve Efficiency of Deep Robotic Grasping,” arXiv:1709.07857 [cs], Sep. 2017(arXiv: 1709.07857)
Article: New Frontier Lin Yi Lun, wearing a star of the original, Li Li, Wang, Wang Feiyue artificial intelligence research: the formula against Network Automation Technology, 2018, 44 (5): 775-792.