Week Ending 1.10.2021
RESEARCH WATCH: 1.10.2021
Over the past week, 692 new papers were published in "Computer Science".
The paper discussed most in the news over the past week was by a team at University of Cambridge: "Hey Alexa what did I just type? Decoding smartphone sounds with a voice assistant" by Almos Zarandy et al (Dec 2020), which was referenced 8 times, including in the article Voice Assistants Can Store And Leak Texts Typed On Smartphones In Proximity in Latest Hacking News. The paper got social media traction with 31 shares. The researchers show that privacy threats go beyond spoken conversations and include sensitive data typed on nearby smartphones. A user, @TwitchiH, tweeted ""...can extract PIN codes and text messages from recordings collected by a voice assistant located up to half a meter away. This shows that remote keyboard-inference attacks are not limited to physical keyboards but extend to virtual keyboards too."".
Leading researcher Kyunghyun Cho (New York University) came out with "Catastrophic Fisher Explosion: Early Phase Fisher Matrix Impacts Generalization".
The paper shared the most on social media this week is by a team at Adobe: "Out of Order: How important is the sequential order of words in a sentence in Natural Language Understanding tasks?" by Thang M. Pham et al (Dec 2020) with 280 shares. @KiddoThe2B (Ilias Chalkidis) tweeted "So instead of burning 🌳 and 💸 in order to hack NLU challenges and do PR, maybe it's a better idea to spend some resources to curate datasets?".
This week was active for "Computer Science - Artificial Intelligence", with 133 new papers.
The paper discussed most in the news over the past week was by a team at Stanford University: "Design Space for Graph Neural Networks" by Jiaxuan You et al (Nov 2020), which was referenced 4 times, including in the article Interesting papers I read from NeurIPS2020 in Towards Data Science. The paper got social media traction with 105 shares. The researchers define and systematically study the architectural design space for GNNs which consists of 315,000 different designs over 32 different predictive tasks. A user, @youjiaxuan, tweeted "We are excited to release #GraphGym, a platform for designing and evaluating #GraphNeuralNetworks. It provides a modularized pipeline, a system for launching thousands of experiments, and more! Code: Paper: #NeurIPS2020 Spotlight".
Leading researcher Quoc V. Le (Google) came out with "AutoDropout: Learning Dropout Patterns to Regularize Deep Networks", which had 23 shares over the past 3 days. The researchers propose to learn the dropout patterns. @hillbig tweeted "AutoDropout learns dropout patterns by RL. These patterns are generated by tiling rectangles and then transform them geometrically. Improve the generalization performance on image recognition, language understanding, and machine translation". This paper was also shared the most on social media with 69 tweets. @hillbig (Daisuke Okanohara) tweeted "AutoDropout learns dropout patterns by RL. These patterns are generated by tiling rectangles and then transform them geometrically. Improve the generalization performance on image recognition, language understanding, and machine translation".
Over the past week, 172 new papers were published in "Computer Science - Computer Vision and Pattern Recognition".
The paper discussed most in the news over the past week was by a team at City University of Hong Kong: "Is a Green Screen Really Necessary for Real-Time Portrait Matting?" by Zhanghan Ke et al (Nov 2020), which was referenced 4 times, including in the article Top 10 Computer Vision Papers 2020 in KDNuggets. The paper got social media traction with 118 shares. A Twitter user, @arxiv_pop, said "2020/11/24 投稿 1位 CV(Computer Vision and Pattern Recognition) Is a Green Screen Really Necessary for Real-Time Human Matting? 6 Tweets 37 Retweets 262 Favorites", while @AkiraTOSEI said "A study of real-time human image cropping. The strategy is to train the model with supervised manner, then self-supervised train with unlabeled data to adjust the model for consistency of three tasks, boundary predictions for low/high resolution mask".
Leading researcher Abhinav Gupta (Carnegie Mellon University) published "Where2Act: From Pixels to Actions for Articulated 3D Objects" The researchers take a step towards that long - term goal -- they extract highly localized actionable information related to elementary actions such as pushing or pulling for articulated objects with movable parts. @DataScienceNIG tweeted "AI figures out how to use real-world objects. Researchers from & have developed a novel method to predict per-pixel actionable information for manipulating articulated 3D objects using the PartNet-Mobility dataset. Paper".
The paper shared the most on social media this week is "VOGUE: Try-On by StyleGAN Interpolation Optimization" by Kathleen M Lewis et al (Jan 2021) with 145 shares. @jacobpedd (Jacob Peddicord) tweeted "This has been one of my fav follows lately. Constantly showing the bleeding edge of ML research. Here is AI powered clothing try on, we will all be using this in a couple years".
This week was active for "Computer Science - Computers and Society", with 31 new papers.
The paper discussed most in the news over the past week was by a team at University of Washington: "Social Media COVID-19 Misinformation Interventions Viewed Positively, But Have Limited Impact" by Christine Geeng et al (Dec 2020), which was referenced 7 times, including in the article Weekend reads: How COVID-19 has changed publications; peer review and women; is ‘manuscript recycling’ OK? in Retraction Watch. The paper got social media traction with 9 shares. On Twitter, @CaulfieldTim commented ""Social Media #COVID19 #Misinformation Interventions Viewed Positively, But Have Limited Impact" cc Platform strategies (general warnings, etc) viewed positively, but specific debunks/corrections likely more effective".
The paper shared the most on social media this week is by a team at IT University of Copenhagen: "The Atlas for the Aspiring Network Scientist" by Michele Coscia (Jan 2021) with 133 shares. @oneofyen (Tzu-Chi Yen) tweeted "This is ... a 760-page book. Whoa".
This week was active for "Computer Science - Human-Computer Interaction", with 27 new papers.
The paper discussed most in the news over the past week was "Can You be More Social? Injecting Politeness and Positivity into Task-Oriented Conversational Agents" by Yi-Chia Wang et al (Dec 2020), which was referenced 2 times, including in the article Social AI? Uber Researchers Propose New Language Model in WebProNews. The paper got social media traction with 15 shares.
This week was very active for "Computer Science - Learning", with 298 new papers.
The paper discussed most in the news over the past week was by a team at Google: "Extracting Training Data from Large Language Models" by Nicholas Carlini et al (Dec 2020), which was referenced 5 times, including in the article Best of arXiv.org for AI, Machine Learning, and Deep Learning – December 2020 in InsideBIGDATA. The paper author, Nicholas Carlini (Google), was quoted saying "We have also worked closely with OpenAI in the analysis of GPT-2". The paper also got the most social media traction with 726 shares. The researchers demonstrate that in such settings, an adversary can perform a training data extraction attack to recover individual training examples by querying the language model. A Twitter user, @shortstein, posted "TL;DR: Snippets of the (public) training data can be extracted from GPT-2. 😮 This is an excellent advertisement for differential privacy research if we want to train on private data. 😀 Blogpost: Paper".
Leading researcher Pieter Abbeel (University of California, Berkeley) published "Reinforcement Learning with Latent Flow".
The paper shared the most on social media this week is "Coding for Distributed Multi-Agent Reinforcement Learning" by Baoqian Wang et al (Jan 2021) with 107 shares. The investigators aim to mitigate straggler effects in synchronous distributed learning for multi - agent reinforcement learning (MARL) problems.
The most influential Twitter user discussing papers is Corey S. Powell who shared "Low Albedo Surfaces of Lava Worlds" by Zahra Essack et al (Aug 2020) and said: "Lava? Exoplanets? Why not have both?".
Over the past week, 12 new papers were published in "Computer Science - Multiagent Systems".
Over the past week, 11 new papers were published in "Computer Science - Neural and Evolutionary Computing".
This week was active for "Computer Science - Robotics", with 57 new papers.
The paper discussed most in the news over the past week was "Chitrakar: Robotic System for Drawing Jordan Curve of Facial Portrait" by Aniruddha Singhal et al (Nov 2020), which was referenced 2 times, including in the article Chitrakar: A system that can transform images of human faces into drawings in Tech Xplore. The paper author, Singhal, was quoted saying "It is an engineering feat to be able to combine state-of-the-art deep-learning methods, image processing and robotics to enable a robotic arm to draw a portrait in its own unique style".
Leading researcher Abhinav Gupta (Carnegie Mellon University) came out with "Where2Act: From Pixels to Actions for Articulated 3D Objects" The investigators take a step towards that long - term goal -- they extract highly localized actionable information related to elementary actions such as pushing or pulling for articulated objects with movable parts. @DataScienceNIG tweeted "AI figures out how to use real-world objects. Researchers from & have developed a novel method to predict per-pixel actionable information for manipulating articulated 3D objects using the PartNet-Mobility dataset. Paper". This paper was also shared the most on social media with 28 tweets. @DataScienceNIG (DataScienceNigeria) tweeted "AI figures out how to use real-world objects. Researchers from & have developed a novel method to predict per-pixel actionable information for manipulating articulated 3D objects using the PartNet-Mobility dataset. Paper".