原始GAN

Goodfellow和Bengio等人發(fā)表在NIPS 2014年的文章Generative adversary network，是生成對(duì)抗網(wǎng)絡(luò)的開(kāi)創(chuàng)文章，論文思想啟發(fā)自博弈論中的二人零和博弈。在二人零和博弈中，兩位博弈方的利益之和為零或一個(gè)常數(shù)，即一方有所得，另一方必有所失。GAN模型中的兩位博弈方分別由生成式模型(generative model)和判別式模型(discriminative model)充當(dāng)。生成模型G捕捉樣本數(shù)據(jù)的分布，判別模型D是一個(gè)二分類器，估計(jì)一個(gè)樣本來(lái)自于訓(xùn)練數(shù)據(jù)（而非生成數(shù)據(jù)）的概率。G和D一般都是非線性映射函數(shù)，例如多層感知機(jī)、卷積神經(jīng)網(wǎng)絡(luò)等。

如圖所示，左圖是一個(gè)判別式模型，當(dāng)輸入訓(xùn)練數(shù)據(jù)x時(shí)，期待輸出高概率（接近1）；右圖下半部分是生成模型，輸入是一些服從某一簡(jiǎn)單分布（例如高斯分布）的隨機(jī)噪聲z，輸出是與訓(xùn)練圖像相同尺寸的生成圖像。向判別模型D輸入生成樣本，對(duì)于D來(lái)說(shuō)期望輸出低概率（判斷為生成樣本），對(duì)于生成模型G來(lái)說(shuō)要盡量欺騙D，使判別模型輸出高概率（誤判為真實(shí)樣本），從而形成競(jìng)爭(zhēng)與對(duì)抗。

GAN.png

GAN優(yōu)勢(shì)很多：根據(jù)實(shí)際的結(jié)果，看上去產(chǎn)生了更好的樣本；GAN能訓(xùn)練任何一種生成器網(wǎng)絡(luò)；GAN不需要設(shè)計(jì)遵循任何種類的因式分解的模型，任何生成器網(wǎng)絡(luò)和任何鑒別器都會(huì)有用；GAN無(wú)需利用馬爾科夫鏈反復(fù)采樣，無(wú)需在學(xué)習(xí)過(guò)程中進(jìn)行推斷，回避了近似計(jì)算棘手的概率的難題。

GAN主要存在的以下問(wèn)題：網(wǎng)絡(luò)難以收斂，目前所有的理論都認(rèn)為GAN應(yīng)該在納什均衡上有很好的表現(xiàn)，但梯度下降只有在凸函數(shù)的情況下才能保證實(shí)現(xiàn)納什均衡。

GAN發(fā)展

一方面GAN的發(fā)展很快，這里只是簡(jiǎn)單粗略將相關(guān)論文分了幾類，歡迎反饋，持續(xù)更新。此外最近ICLR 2017 在進(jìn)行Open Review，可以關(guān)注下ICLR 2017 Conference Track，也有相應(yīng)論文筆記分享ICLR 2017 | GAN Missing Modes 和 GAN

GAN從2014年到現(xiàn)在發(fā)展很快，特別是最近ICLR 2016/2017關(guān)于GAN的論文很多，GAN現(xiàn)在有很多問(wèn)題還有到解決，潛力很大。總體可以將已有的GANs論文分為以下幾類

1. GAN Theory
2. GAN in Semi-supervised
3. Muti-GAN
4. GAN with other Generative model
5. GAN with RNN
6. GAN in Application

GAN Theory

此類關(guān)注與無(wú)監(jiān)督GAN本身原理的研究：比較兩個(gè)分布的距離；用DL的一些方法讓GAN快速收斂等等。相關(guān)論文有：

• GAN: Goodfellow, Ian, et al. "Generative adversarial nets." Advances in Neural Information Processing Systems. 2014.
• LAPGAN: Denton, Emily L., Soumith Chintala, and Rob Fergus. "Deep Generative Image Models using a Laplacian Pyramid of Adversarial Networks." Advances in neural information processing systems. 2015.
• DCGAN: Radford, Alec, Luke Metz, and Soumith Chintala. "Unsupervised representation learning with deep convolutional generative adversarial networks." arXiv preprint arXiv:1511.06434 (2015).
• Improved GAN: Salimans, Tim, et al. "Improved techniques for training gans." arXiv preprint arXiv:1606.03498 (2016).
• InfoGAN: Chen, Xi, et al. "Infogan: Interpretable representation learning by information maximizing generative adversarial nets." arXiv preprint arXiv:1606.03657(2016).**
• EnergyGAN: Zhao, Junbo, Michael Mathieu, and Yann LeCun. "Energy-based Generative Adversarial Network." arXiv preprint arXiv:1609.03126 (2016).
• Creswell, Antonia, and Anil A. Bharath. "Task Specific Adversarial Cost Function." arXiv preprint arXiv:1609.08661 (2016).
• f-GAN: Nowozin, Sebastian, Botond Cseke, and Ryota Tomioka. "f-GAN: Training Generative Neural Samplers using Variational Divergence Minimization." arXiv preprint arXiv:1606.00709 (2016).
• Unrolled Generative Adversarial Networks, ICLR 2017 Open Review
• Improving Generative Adversarial Networks with Denoising Feature Matching, ICLR 2017 Open Review
• Mode Regularized Generative Adversarial Networks, ICLR 2017 Open Review
• b-GAN: Unified Framework of Generative Adversarial Networks, ICLR 2017 Open Review
• Mohamed, Shakir, and Balaji Lakshminarayanan. "Learning in Implicit Generative Models." arXiv preprint arXiv:1610.03483 (2016).

GAN in Semi-supervised

此類研究將GAN用于半監(jiān)督學(xué)習(xí)，相關(guān)論文有：

• Springenberg, Jost Tobias. "Unsupervised and Semi-supervised Learning with Categorical Generative Adversarial Networks." arXiv preprint arXiv:1511.06390 (2015).
• Odena, Augustus. "Semi-Supervised Learning with Generative Adversarial Networks." arXiv preprint arXiv:1606.01583 (2016).

Muti-GAN

此類研究將多個(gè)GAN進(jìn)行組合，相關(guān)論文有：

• CoupledGAN: Liu, Ming-Yu, and Oncel Tuzel. "Coupled Generative Adversarial Networks." arXiv preprint arXiv:1606.07536 (2016).
• Wang, Xiaolong, and Abhinav Gupta. "Generative Image Modeling using Style and Structure Adversarial Networks." arXiv preprint arXiv:1603.05631(2016).
• Generative Adversarial Parallelization, ICLR 2017 Open Review
• LR-GAN: Layered Recursive Generative Adversarial Networks for Image Generation, ICLR 2017 Open Review

GAN with other Generative model

此類研究將GAN與其他生成模型組合，相關(guān)論文有：

• Dosovitskiy, Alexey, and Thomas Brox. "Generating images with perceptual similarity metrics based on deep networks." arXiv preprint arXiv:1602.02644(2016).
• Larsen, Anders Boesen Lindbo, S?ren Kaae S?nderby, and Ole Winther. "Autoencoding beyond pixels using a learned similarity metric." arXiv preprint arXiv:1512.09300 (2015).
• Theis, Lucas, and Matthias Bethge. "Generative image modeling using spatial lstms." Advances in Neural Information Processing Systems. 2015.

GAN with RNN

此類研究將GAN與RNN結(jié)合（也以參考Pixel RNN），相關(guān)論文有：

• Im, Daniel Jiwoong, et al. "Generating images with recurrent adversarial networks." arXiv preprint arXiv:1602.05110 (2016).
• Kwak, Hanock, and Byoung-Tak Zhang. "Generating Images Part by Part with Composite Generative Adversarial Networks." arXiv preprint arXiv:1607.05387 (2016).
• Yu, Lantao, et al. "SeqGAN: Sequence Generative Adversarial Nets with Policy Gradient." arXiv preprint arXiv:1609.05473 (2016).

GAN in Application

此類研究將GAN的實(shí)際運(yùn)用（不包括圖像生成），相關(guān)論文有：

• Zhu, Jun-Yan, et al. "Generative visual manipulation on the natural image manifold." European Conference on Computer Vision. Springer International Publishing, 2016.
• Creswell, Antonia, and Anil Anthony Bharath. "Adversarial Training For Sketch Retrieval." European Conference on Computer Vision. Springer International Publishing, 2016.
• Reed, Scott, et al. "Generative adversarial text to image synthesis." arXiv preprint arXiv:1605.05396 (2016).
• Ravanbakhsh, Siamak, et al. "Enabling Dark Energy Science with Deep Generative Models of Galaxy Images." arXiv preprint arXiv:1609.05796(2016).
• Abadi, Martín, and David G. Andersen. "Learning to Protect Communications with Adversarial Neural Cryptography." arXiv preprint arXiv:1610.06918(2016).
• Odena, Augustus, Christopher Olah, and Jonathon Shlens. "Conditional Image Synthesis With Auxiliary Classifier GANs." arXiv preprint arXiv:1610.09585 (2016).
• Ledig, Christian, et al. "Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network." arXiv preprint arXiv:1609.04802 (2016).
• Nguyen, Anh, et al. "Synthesizing the preferred inputs for neurons in neural networks via deep generator networks." arXiv preprint arXiv:1605.09304(2016).