问:在训练GAN方面似乎有两三个阵营(camp),你和在OpenAI、谷歌的人,Mescheder, Sebastian Nowozin和微软研究院的人,以及其他。在这些阵营中你有看到什么相似点吗?
Ian Goodfellow:实际上有更多阵营。FAIR/NYU也是重镇,实际上是FAIR/NYU最先把GAN带进了我们的视野(LAPGAN)。那是第一次GAN生成了逼真的高清图像,也是第一次GAN得到媒体曝光,等等。另外一个重要阵营是伯克利+英伟达,他们专注超高清逼真图像和视频,无监督翻译,等等。我不太清楚你用“阵营”(camps)想表达什么,如果是说用不同的思路和方法在研究GAN,那么确实如此。但“阵营”一般多指相互敌对的团队,在这里并不适用。
Ian Goodfellow:我不认为这个问题有点偏,因为GAN的判别器需要具有鲁棒性,应对生成器的对抗攻击。同时,我确实认为我们当前使用的神经元,很难在应对对抗样本方面具有鲁棒性。但我并不认为这是唯一有问题的地方。一些最新的工作,比如(https://arxiv.org/abs/1801.02774)表明,要让图像分类安全可靠,我们真的需要从根本上重新考虑我们使用的策略,而不仅仅是训练的模型。
Ian Goodfellow:这取决于你想用GAN来做什么。如果你想将其用于半监督式学习,请使用测试集精度作为评估指标,如果你想使用它来生成高质量的图像(例如超分辨率),那么可能需要使用人类评分员来评估。如果你只是想得到一个通用的自动化的质量得分,那么我认为Frechet Inception Distance(https://arxiv.org/abs/1706.08500)可能是最好的,尤其是对于class-specific(https://arxiv.org/pdf/1802.05957.pdf)的模型来说。这些指标本身现在仍是一个正在进行的重要的研究领域。
Ian Goodfellow:我在Fermat’s Library这里做AMA,就是为了推广这些工具,我认为它们很重要,也值得更多人重视。ArXiv现在基本成了绕过同行评议的捷径,让深度学习论文的信噪比骤降。现在仍然有很多优秀的深度学习研究在进行,但同样也存在大量的低质量工作。如今就连最好的那些工作也是好坏掺半——有很好的想法,但却用宣传推广的套路在写,跟其他工作的比较也不准确,等等。这都是因为这些论文没有经过同行评议。
Ian Goodfellow:我猜应该有,但我不知道具体的。文档是离散token,比如文字、单词,而GAN需要计算生成器的输出的梯度,因此在有连续输出的地方才能很好的工作。或许你可以用对抗自编码器,或者AVB,在这里生成器实际上是编码器,能够输出连续的代码。这对于文档建模有很大帮助,因为能对topics给出一个分布式表示。
问:GAN在基因组学里有什么应用?
Ian Goodfellow:我不太了解基因组学,但我认为GAN这类的模型可以用于半监督学习。我想在基因组学中,未标记的数据应该比有标记的更多,利用大量的未标记数据从少量标记数据中学习的话,半监督学习会很有帮助。这方面Tim Salimans提出了一些方法,在MNIST、SVHN等基准上特别好用:https://arxiv.org/abs/1606.03498
"Defending Against Adversarial Examples". NIPS 2017 Workshop on Machine Learning and Security. [slides(pdf)] [slides(key)]
"Thermometer Encoding: One hot way to resist adversarial examples," 2017-11-15, Stanford University [slides(pdf)] [slides(key)]
"Adversarial Examples and Adversarial Training," 2017-05-30, CS231n, Stanford University [slides(pdf)] [slides(key)]
"Adversarial Examples and Adversarial Training," 2017-01-17, Security Seminar, Stanford University [slides(pdf)] [slides(key)]
"Adversarial Examples and Adversarial Training," 2016-12-9, NIPS Workshop on Reliable ML in the Wild [slides(pdf)] [slides(key)] [video(wmv)]
"Adversarial Examples and Adversarial Training," presentation at Uber, October 2016. [slides(pdf)]
"Physical Adversarial Examples," presentation and live demo at GeekPwn 2016 with Alex Kurakan. [slides(pdf)]
"Adversarial Examples and Adversarial Training," guest lecture for CS 294-131 at UC Berkeley. [slides(pdf)] [slides(key)] [video(youtube)]
"Exploring vision-based security challenges for AI-driven scene understanding," joint presentation with Nicolas Papernot at AutoSens, September 2016, in Brussels. Access to the slides and video may be purchased at the conference website. They will be freely available after six months.
"Adversarial Examples and Adversarial Training" at HORSE 2016. [slides(pdf)] [youtube]
"Adversarial Examples and Adversarial Training" at San Francisco AI Meetup, 2016. [slides(pdf)]
"Adversarial Examples and Adversarial Training" at Quora, Mountain View, 2016. [slides(pdf)]
"Adversarial Examples" at the Montreal Deep Learning Summer School, 2015. [slides(pdf)] [video]
"Do statistical models understand the world?" Big Tech Day, Munich, 2015. [youtube]
"Adversarial Examples" Re-Work Deep Learning Summit, 2015. [youtube]
Generative Adversarial Networks
"Overcoming Limited Data with GANs". NIPS 2017 Workshop on Limited Labeled Data. [slides(pdf)] [slides(key)]
"Bridging theory and practice of GANs". NIPS 2017 Workshop on Bridging Theory and Practice of Deep Learning. [slides(pdf)] [slides(key)]
"GANs for Creativity and Design". NIPS 2017 Workshop on Creativity and Design. [slides(pdf)] [slides(key)]
"Giving artificial intelligence imagination using game theory". 35 under 35 talk at EmTech 2017. [slides(pdf)][slides(key)]
"Generative Adversarial Networks". Introduction to ICCV Tutorial on Generative Adversarial Networks, 2017. [slides(pdf)] [slides(key)]
"Generative Adversarial Networks," a guest lecture for John Canny's COMPSCI 294 at UC Berkeley. Oct 2016. [slides(keynote)] [slides(pdf)] [youtube]
"Generative Adversarial Networks" at AI With the Best (online conference), September 2016. [slides(pdf)]
"Generative Adversarial Networks" keynote at MLSLP, September 2016, San Francisco. [slides]
"Generative Adversarial Networks" at Berkeley AI Lab, August 2016. [slides(pdf)]
"Generative Adversarial Networks" at NVIDIA GTC, April 2016. [slides(pdf)][video]
"Generative Adversarial Networks" at ICML Deep Learning Workshop, Lille, 2015. [slides(pdf)] [video]
"Generative Adversarial Networks" at NIPS Workshop on Perturbation, Optimization, and Statistics, Montreal, 2014. [slides(pdf)]
Other Subjects
"Adversarial Robustness for Aligned AI". NIPS 2017 Workshop on Aligned AI. [slides(pdf)] [slides(key)]
"Defense Against the Dark Arts: Machine Learning Security and Privacy," BayLearn, 2017-10-19. [slides(pdf)][video(youtube)]
"Adversarial Machine Learning for Security and Privacy," Army Research Organization workshop, Stanford, 2017-09-14. [slides(pdf)]
"Generative Models I," 2017-06-27, MILA Deep Learning Summer School. [slides(pdf)] [slides(key)]
"Adversarial Approaches to Bayesian Learning and Bayesian Approaches to Adversarial Robustness," 2016-12-10, NIPS Workshop on Bayesian Deep Learning [slides(pdf)] [slides(key)]
"Design Philosophy of Optimization for Deep Learning" at Stanford CS department, March 2016. [slides(pdf)]
"Tutorial on Optimization for Deep Networks" Re-Work Deep Learning Summit, 2016. [slides(keynote)] [slides(pdf)]
"Tutorial on Neural Network Optimization Problems" at the Montreal Deep Learning Summer School, 2015. [slides(pdf)][video]
"Practical Methodology for Deploying Machine Learning" Learn AI With the Best, 2015. [slides(pdf)] [youtube]
Contributed Talks
"Qualitatively characterizing neural network optimization problems" at ICLR 2015. [slides(pdf)]
"Multi-digit Number Recognition from Street View Imagery using Deep Convolutional Neural Networks" with Yaroslav Bulatov and Julian Ibarz at ICLR 2014. [youtube]
"Maxout Networks" at ICML 2013. [video]
"Joint Training Deep Boltzmann Machines for Classification" at ICLR 2013 (workshop track). [video]
Miscellaneous
I've made several presentations for the Deep Learning textbook, and presented some of them at a study group for the book.
PhD thesis defense. [youtube] [slides]
Ian Goodfellow GAN资料地址:http://www.iangoodfellow.com/slides
Ian Goodfellow 关于GAN的最新AMA地址:https://fermatslibrary.com/arxiv_comments?url=https%3A%2F%2Farxiv.org%2Fpdf%2F1406.2661.pdf