CS graduate student Allan Jabri has been named a 2018 Paul & Daisy Soros Fellow.   Soros Fellowships are awarded to outstanding immigrants and children of immigrants from across the globe who are pursuing graduate school in the United States.  Recipients are chosen for their potential to make significant contributions to US society, culture, or their academic fields, and will receive up to $90K in funding over two years.  Jabri was born in Australia to parents from China and Lebanon and was raised in the US.   He received his B.S. at Princeton where his thesis focused on probabilistic methods for egocentric scene understanding, and worked as a research engineer at Facebook AI Research in New York before joining Berkeley AI Research (BAIR).  He  is interested in problems related to self-supervised learning, continual learning, intrinsic motivation, and embodied cognition. His long-term goal is to build learning algorithms that allow machines to autonomously acquire visual and sensorimotor common sense. During his time at Berkeley, he also hopes to mentor students, contribute to open source code projects, and develop a more interdisciplinary perspective on AI.

EE graduate student Stephen Tu (advisor: Ben Recht) has been awarded a 2018 Google Fellowship.  Google Fellowships are presented to exemplary PhD students in computer science and related areas to acknowledge contributions to their chosen fields and provide funding for their education and research. Tu's current research interests "lie somewhere in the intersection of machine learning and optimization" although he previously worked on multicore databases and encrypted query processing.  Tu graduated with a CS B.A./ME B.S. from Berkeley in 2011 before earning an EECS S.M. from MIT in 2014.

Graduate student Xue Bin “Jason” Peng (advisors Pieter Abbeel and Sergey Levine) has made a major advance in realistic computer animation using deep reinforcement learning to recreate natural motions, even for acrobatic feats like break dancing and martial arts. The simulated characters can also respond naturally to changes in the environment, such as recovering from tripping or being pelted by projectiles.  “We developed more capable agents that behave in a natural manner,” Peng said. “If you compare our results to motion-capture recorded from humans, we are getting to the point where it is pretty difficult to distinguish the two, to tell what is simulation and what is real. We’re moving toward a virtual stuntman.”  Peng will present his paper at the 2018 SIGGRAPH conference in August.

Prof. Ali Javey,  postdoc Der-Hsien Lien, and graduate students Matin Amani and Sujay Desai have built a bright-light emitting device that is millimeters wide and fully transparent when turned off.  The light emitting material in this device is a monolayer semiconductor, which is just three atoms thick.  It opens the door to invisible displays on walls and windows – displays that would be bright when turned on but see-through when turned off — or in futuristic applications such as light-emitting tattoos.  “The materials are so thin and flexible that the device can be made transparent and can conform to curved surfaces,” said  Lien. Their research was published in the journal Nature Communications on March 26.

Undergraduate students Ganeshkumar Ashokavardhanan (EECS  + Business M.E.T.),  Naomi Jung (CS BA), and Louie McConnell (EECS + Business M.E.T.) have been selected to participate in the 2018 KPCB Engineering Fellows Program, named one of the top 5 internship programs by Vault.  Over the course of a summer, KPCB Engineering Fellows join portfolio companies, where they develop their technical skills and are each mentored by an executive within the company. It offers students an opportunity to gain significant work experience at Silicon Valley startups, collaborating on unique and challenging technical problems.

A paper released on arXiv last week by a team of researchers including Prof. Dawn Song and Ph.D. student Nicholas Carlini (B.A. CS/Math '13), reveals just how vulnerable deep learning is to information leakage.  The researchers labelled the problem “unintended memorization” and explained it happens if miscreants can access to the model’s code and apply a variety of search algorithms. That's not an unrealistic scenario considering the code for many models are available online, and it means that text messages, location histories, emails or medical data can be leaked.  The team doesn't “really know why neural networks memorize these secrets right now, ” Carlini says.  “At least in part, it is a direct response to the fact that we train neural networks by repeatedly showing them the same training inputs over and over and asking them to remember these facts."   The best way to avoid all problems is to never feed secrets as training data. But if it’s unavoidable then developers will have to apply differentially private learning mechanisms, to bolster security, Carlini concluded.

CS graduate student Ling-Qi Yan (advisors: Ravi Ramamoorthi/Ren Ng) and researchers at U.C. San Diego are the subject of an article in TechXplore titled "Scientists improve computer rendering of animal fur."  He is part of a team that developed a method for dramatically improving the way computers simulate fur, and more specifically, the way light bounces within an animal's pelt.  The researchers are using a neural network to apply the properties of a concept called subsurface scattering to quickly approximate how light bounces around fur fibers.  The neural network only needs to be trained with one scene before it can apply subsurface scattering to all the different scenes with which it is presented. This results in simulations running 10 times faster than current state of the art.  "We are converting the properties of subsurface scattering to fur fibers," said Yan. "There is no explicit physical or mathematical way to make this conversion. So we needed to use a neural network to connect these two different worlds."  The researchers recently presented their findings at the SIGGRAPH Asia conference in Thailand.

For the second year running, UC Berkeley has won the Fiesta Bowl Overwatch Collegiate National Championship, sweeping UC Irvine 3-0.  CS majors Kevin "SlurpeeThief" Royston and Gandira “Syeikh” Prahandika were on the Berkeley team, which battled in front of a sold-out crowd at the game in Tempe, Arizona, the first partnership between a collegiate bowl game and eSports tournament.  Royston, who was profiled in on Overwatch Wire article before the tournament, is in the top 1% of all players worldwide, has had a peak skill rating of 4626, and was on the winning team last year.  He notes that the balance between esports and school has been a tough road.  “It’s definitely rough, this week is the first time I had a homework assignment slip because of Overwatch. (Laughs) Hear me out! It’s for Machine Learning which is the hardest class at Berkeley. We had to travel (for the Fiesta Bowl) and I worked sixteen hours on an assignment and didn’t even get halfway through it.”  Overwatch is a team-based shooter game that was created by Blizzard Entertainment and released in 2016.  Teams are made up of six players, who select different characters, known in the game as “heroes,” to complete different objectives. The Berkeley team won $42,000 of the $100,000 total in scholarships and other prizes.

Prof. Dawn Song and graduate student Noah Johnson are taking a new approach to enable organizations to follow tight data security and privacy policies while enabling flexible data analysis, as well as machine learning for analysts.  Working with Uber, they tested their system using a dataset of 8 million queries written by the company’s data analysts. The system is currently being integrated into Uber’s internal data analytics platform.  With help from the Signatures Innovation Fellows program, they are advancing the system to provide the same level of security and flexibility for a broad range of data analysis and machine learning, whether needed in basic and medical research or business analytics.

Prof. Alexandre Bayen and PhD student Pulkit Agrawal developed a computer vision-based system to help memory care centers monitor patient falls and to reduce them where possible.  State regulations require an MRI of the head any time a patient suffers an unwitnessed fall, and about a fourth of all Alzheimer’s-related hospital visits are triggered by a fall. With five million Americans currently living with Alzheimer’s, the task of preventing, tracking and treating fall-related injuries has become daunting and costly, with more than a $5 billion annual cost to medicare--and the number of people with Alzheimer’s is expected to double in the next 15 years.   A system capable of detecting falls by autonomously monitoring patients and sending therapists video clips could improve the monitoring process immensely  “There are no effective drugs yet to treat Alzheimer’s,” Agarwal says. “Until we have them, we have to help patients where they are. Developing computer vision systems to detect falls and fall vulnerability seemed like a good way to improve healthcare for a growing patient population.”