Eye On AI

View Original

Week Ending 05.12.19

Craig Smith

RESEARCH WATCH: 05.12.19

Over the past week, 209 new papers were published in "Computer Science".

Over the past week, 92 new papers were published in "Computer Science - Artificial Intelligence".

  • The paper discussed most in the news over the past week was by a team at University of California, Berkeley: "Semantic Image Synthesis with Spatially-Adaptive Normalization" by Taesung Park et al (Mar 2019), which was referenced 21 times, including in the article Implementing SPADE using fastai in Medium.com. The paper also got the most social media traction with 775 shares. A Twitter user, @spiltlens, said "give this tech 10 more years and all these movies that are just shots of different handwritten letters over and over again arent even gonna be recognizable as cinema", while @avsa commented "Source for that (and many other GAN algorithms) are flickr cc photosets. Flickr is like that whale that, for many years after its death, it still feeds a thriving ecosystem on its remains".

  • Leading researcher Quoc V. Le (Google) came out with "The Effect of Network Width on Stochastic Gradient Descent and Generalization: an Empirical Study"@arankomatsuzaki tweeted "The Effect of Network Width on Stochastic Gradient Descent and Generalization: an Empirical Study: test accuracy strongly depends on the network width, which has a strong correlation with the normalized noise scale".

  • The paper shared the most on social media this week is "Few-Shot Unsupervised Image-to-Image Translation" by Ming-Yu Liu et al (May 2019) with 321 shares. @DanilBaibak (Danylo Baibak) tweeted "Few-Shot Unsupervised Image-to-Image Translation - one more interesting paper from #NVidia #NeuralNetwork #DeepLearning".

Over the past week, 187 new papers were published in "Computer Science - Computer Vision and Pattern Recognition".

  • The paper discussed most in the news over the past week was "Fooling automated surveillance cameras: adversarial patches to attack person detection" by Simen Thys et al (Apr 2019), which was referenced 36 times, including in the article Researchers hack AI video analytics with color printout in Security Dealer. The paper author, Wiebe Van Ranst, was quoted saying "The idea behind this work is to be able to circumvent security systems that use a person detector to generate an alarm when a person enters the view of a camera". The paper also got the most social media traction with 13980 shares. On Twitter, @pwang observed "This is going to be a major Tshirt trend over the next couple of years: Innocent-looking shirts with various patterns that are specifically designed to trick neural networks. Adversarial hats will also be a thing, for face detectors".

  • Leading researcher Quoc V. Le (Google) published "Searching for MobileNetV3", which had 21 shares over the past 3 days. The investigators start the exploration of how automated search algorithms and network design can work together to harness complementary approaches improving the overall state of the art. @tobyminion tweeted "MobileNetV3 is coming 😶😶😶".

  • The paper shared the most on social media this week is by a team at Massachusetts Institute of Technology: "Adversarial Examples Are Not Bugs, They Are Features" by Andrew Ilyas et al (May 2019) with 504 shares. @SashaVNovikov (Alexander Novikov) tweeted "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".

Over the past week, 16 new papers were published in "Computer Science - Computers and Society".

  • The paper discussed most in the news over the past week was "Discrimination through optimization: How Facebooks ad delivery can lead to skewed outcomes" by Muhammad Ali et al (Apr 2019), which was referenced 52 times, including in the article All the Ways Hiring Algorithms Can Introduce Bias in Harvard Business Review. The paper author, Alan Mislove (Computer science professor at Northeastern University), was quoted saying "All advertising is based on auctions all over the web, and I don’t know how you fix that without just saying we don’t have those kinds of ads". The paper also got the most social media traction with 1319 shares. The investigators demonstrate that such skewed delivery occurs on Facebook, due to market and financial optimization effects as well as the platforms own predictions about the relevance of ads to different groups of users. A Twitter user, @TimKarr, observed "April 4, 2019: Academic paper published analyzing FB's algorithms to find they're built using historically discriminatory data. The algorithms deliver results biased against people based on race & gender, & perpetuate discrimination in advertising".

This week was active for "Computer Science - Human-Computer Interaction", with 28 new papers.

Over the past week, 175 new papers were published in "Computer Science - Learning".

  • The paper discussed most in the news over the past week was by a team at University of California, Berkeley: "Semantic Image Synthesis with Spatially-Adaptive Normalization" by Taesung Park et al (Mar 2019), which was referenced 21 times, including in the article Implementing SPADE using fastai in Medium.com. The paper got social media traction with 775 shares. On Twitter, @spiltlens posted "give this tech 10 more years and all these movies that are just shots of different handwritten letters over and over again arent even gonna be recognizable as cinema", while @avsa said "Source for that (and many other GAN algorithms) are flickr cc photosets. Flickr is like that whale that, for many years after its death, it still feeds a thriving ecosystem on its remains".

  • Leading researcher Oriol Vinyals (DeepMind) published "REGAL: Transfer Learning For Fast Optimization of Computation Graphs".

  • The paper shared the most on social media this week is by a team at Massachusetts Institute of Technology: "Adversarial Examples Are Not Bugs, They Are Features" by Andrew Ilyas et al (May 2019) with 504 shares. @SashaVNovikov (Alexander Novikov) tweeted "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".

Over the past week, ten new papers were published in "Computer Science - Multiagent Systems".

Over the past week, 24 new papers were published in "Computer Science - Neural and Evolutionary Computing".

  • The paper discussed most in the news over the past week was by a team at Stanford University: "Wave Physics as an Analog Recurrent Neural Network" by Tyler W. Hughes et al (Apr 2019), which was referenced 6 times, including in the article Putin Will Put Russia Behind an Web Curtain – NEWPAPER24 in Newpaper24. The paper also got the most social media traction with 269 shares. The researchers identify a mapping between the dynamics of wave physics, and the computation in recurrent neural networks. A user, @KordingLab, tweeted "This is the beginning of practically conceptualizing physics as computation", while @twhughes91 said "Episode 153 (Possible Minds) of Making Sense podcast is super interesting! Our work on analog machine learning with wave-based physics is one answer to the question "What are some examples of analog computing in 2019"".

Over the past week, 43 new papers were published in "Computer Science - Robotics".

EYE ON A.I. GETS READERS UP TO DATE ON THE LATEST FUNDING NEWS AND RELATED ISSUES. SUBSCRIBE FOR THE WEEKLY NEWSLETTER.