Papers of the day   All papers

Adversarial Examples Are Not Bugs, They Are Features

Comments

Chris Olah: “Adversarial Examples Are Not Bugs, They Are Features” by Ilyas et al is pretty interesting. 📝Paper: https://arxiv.org/pdf/1905.02175.pdf 💻Blog: http://gradientscience.org/adv/ Some quick notes below.

3 replies, 712 likes


Ilya Sutskever: https://arxiv.org/abs/1905.02175: a strange imagenet-like dataset with very wrong-looking labels, yet a model trained on it does totally well on the normal validation set. It's a crime against ML!

10 replies, 526 likes


Lilian Weng: Two most interesting papers I’ve found recently: “the lottery ticket hypothesis” https://openreview.net/forum?id=rJl-b3RcF7 (probably already very famous) and “adversarial examples are not bugs but features” https://arxiv.org/abs/1905.02175

7 replies, 263 likes


Louise Matsakis: Researchers from MIT now think adversarial examples aren’t AI “hallucinations” after all. The classifier is just “seeing” things that humans can’t. It’s really interesting work! https://www.wired.com/story/adversarial-examples-ai-may-not-hallucinate/

2 replies, 121 likes


Wojciech Zaremba: Adversarial examples are great features. There is nothing wrong with them. We are just blind to them. Wow. https://arxiv.org/pdf/1905.02175.pdf

1 replies, 89 likes


Aleksander Madry: A great piece by @lmatsakis about our recent work on how adversarial examples are actually helpful features! To read more see: http://gradientscience.org/adv and the paper is here: https://arxiv.org/abs/1905.02175.

2 replies, 36 likes


Hamid Eghbal-zadeh: So https://arxiv.org/abs/1905.02175 says we can use adversarial examples to train models that generalise. But https://openreview.net/forum?id=rJMw747l_4 says if you train a model with samples from a GAN, it does not generalise, unless you mix it up with real samples!🤔

2 replies, 29 likes


Shreya Shankar: does this stuff interest you? some good papers (in my opinion): * Motivating the Rules of the Game for Adversarial Example Research: https://arxiv.org/abs/1807.06732 (Gilmer et al. 2018) * Adversarial Examples Are Not Bugs, They Are Features: https://arxiv.org/abs/1905.02175 (Ilyas et al. 2019)

1 replies, 28 likes


Thomas Lahore: Adversarial Examples Are Not Bugs, They Are Features "adversarial examples can be directly attributed to the presence of non-robust features ... patterns in the data distribution that are highly predictive, yet brittle and incomprehensible to humans" https://arxiv.org/abs/1905.02175

0 replies, 24 likes


Daisuke Okanohara: Adversarial examples are not the result of artifacts or overfitting. It is because the learned classifier captures "non-robust" features, which are not incomprehensible by humans but are actually predictive and generalizable. https://arxiv.org/abs/1905.02175

0 replies, 11 likes


Adam J Calhoun: What if adversarial examples exist because they help the network generalize? And it is simply our puny human minds that aren't able to understand that? https://arxiv.org/abs/1905.02175 https://t.co/JPvNWCjqZG

0 replies, 11 likes


午後のarXiv: "Adversarial Examples Are Not Bugs, They Are Features", Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan En… https://arxiv.org/abs/1905.02175

0 replies, 7 likes


Sai Prasanna: @zacharylipton Adversarial Examples Are Not Bugs, They Are Features https://arxiv.org/abs/1905.02175

1 replies, 6 likes


Alexander Novikov: Cool! Like furry ear is a feature which can be used to detect cats, adversarial perturbations are features of natural images which can be used to correctly classify both train and test data, except humans don't see it. So adversarial perturbations are human's bugs, not model's.

0 replies, 6 likes


Stefanie Sirén-Heikel: What if adversarial examples aren't bugs, but features? Check out the story on how we might actually really misunderstand how machine learning systems work: https://www.wired.com/story/adversarial-examples-ai-may-not-hallucinate/ and read the paper here: https://arxiv.org/abs/1905.02175

1 replies, 4 likes


Jason Taylor @NeurIPS2019: Adversarial Examples Are Not Bugs, They Are Features https://arxiv.org/abs/1905.02175 main idea: adversarial examples arise from overfitting on non-robust features. Also shows that adversarial training with flipped labels generalizes to the test set with correct labels

1 replies, 3 likes


Bobby Filar: "Adversarial Examples Are Not Bugs, They Are Features" by @andrew_ilyas, @tsiprasd et al. https://arxiv.org/pdf/1905.02175.pdf

0 replies, 3 likes


@reiver ⊼ (Charles Iliya Krempeaux): "Adversarial Examples Are Not Bugs, They Are Features" by Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, Aleksander Madry https://arxiv.org/abs/1905.02175 (machine learning) H/T @QEDanMazur

0 replies, 2 likes


Eric Silverman: This is a really cool paper! Adversarial examples are getting a lot of attention in deep learning research and this paper offers some real insight into the issue

0 replies, 2 likes


Benjamin Singleton: Adversarial Examples Are Not Bugs, They Are Features #DataScience #BigData https://arxiv.org/abs/1905.02175

0 replies, 1 likes


Hacker News: Adversarial Examples Are Not Bugs, They Are Features: https://arxiv.org/abs/1905.02175 Comments: https://news.ycombinator.com/item?id=19863913

0 replies, 1 likes


Hacker News 20: Adversarial Examples Are Not Bugs, They Are Features https://arxiv.org/abs/1905.02175 (http://bit.ly/2YimheO)

0 replies, 1 likes


Trustworthy ML: Adversarial Examples Are Not Bugs, They Are Features(https://arxiv.org/pdf/1905.02175.pdf)---This paper demonstrates that the non-robust features in the dataset account for adversarial examples, and training the model on robust features w/o adversarial training gives a robust model. 1/2

1 replies, 1 likes


HQME: @CBSarrangesme @watch_m3n hmm i'd need to listen to those samples, easy w the data audio -> melspectrograms, encodes frequency decibels and times in a picture https://arxiv.org/pdf/1905.02175.pdf

5 replies, 1 likes


Wilfred Hughes: Really interesting paper exploring adversarial inputs to ML models: https://arxiv.org/abs/1905.02175 They conclude: * It's a property of the input data, not the training * You can even train a model on non-robust features and obtain a model that works well on the original input data!

0 replies, 1 likes


Covered Dish People: #1yrago Towards a method for fixing machine learning's persistent and catastrophic blind spots https://arxiv.org/pdf/1905.02175.pdf #1yrago Americans with diabetes are forming caravans to buy Canadian insulin at 90% off https://www.cbc.ca/news/canada/nova-scotia/americans-diabetes-cross-canada-border-insulin-1.5125988 13/ https://t.co/qKBCdiFOWZ

1 replies, 1 likes


Chaz Firestone: @IdanAsherBlank @fierycushman but i think that's not quite right. illusions reflect our visual system doing what it's supposed to; a mind that didn't have them would probably end up with some other problems. most people don't think of adversarial images that way (though see https://arxiv.org/abs/1905.02175). 2/n

1 replies, 0 likes


Content

Found on May 09 2019 at https://arxiv.org/pdf/1905.02175.pdf

PDF content of a computer science paper: Adversarial Examples Are Not Bugs, They Are Features