Dr. T's AI brief
Daniel Tauritz
New AI System To Help NASA’s Perseverance Rover Land On Mars
The Washington Post 
(2/16, Achenbach, Guarino, Davenport) reports that the upcoming Thursday land of NASA’s Perseverance rover will be “one of the hardest technological feats human beings have ever attempted,” with the Post comparing it to “throwing a dart from the White House and scoring a bull’s eye in Dallas.” NASA Perseverance rover mission Deputy Project Manager Matt Wallace said, “Success depends on everything going right, down to fractions of a second. There’s no go-back, no retry.” NASA’s spacecraft will use “a new system, packed with artificial intelligence, to scan the terrain and match it with maps of Mars to try to pick out a landing spot.” However, unlike previous terrain chosen for Mars landings, Jezero Crater is “crammed with boulders, gullies, cliffs and” the remnants of an ancient river delta.
CBS News (2/16, Harwood) reports Jezero Crater, which hosted an ancient lake fed by a river, “represents a golden opportunity to find out” whether Mars ever hosted life. Perseverance “will collect promising rock and soil samples, seal them in small lipstick-size containers and deposit them on the surface in carefully identified caches.” Lori Glaze, director of planetary science at NASA Headquarters, said, “In 2026, a fetch rover will be launched to collect those samples and bring them to a rocket that will launch them into orbit around Mars. Another orbiter will rendezvous and capture those samples for safe delivery to Earth.”
IBM Showcases New Energy-efficient AI Accelerator Chip
ZDNet (2/18, Leprince-Ringuet) reports that IBM researchers “have designed what they claim to be the world’s first AI accelerator chip that is built on high-performance seven-nanometer technology, while also achieving high levels of energy efficiency.” IBM research staff members Ankur Agrawal and Kailash Gopalakrishnan “unveiled the four-core chip at the International Solid-State Circuits Virtual Conference this month, and have disclosed more details about the technology in a recent blog post.” Although still at the research stage, “the accelerator chip is expected to be capable of supporting various AI models and of achieving ‘leading’ edge power efficiency.”
Google Restructures Responsible AI Teams
Bloomberg (2/18, Grant, Bass) reports Google “restructured its responsible artificial intelligence efforts to centralize teams under a single executive, Marian Croak,” in an effort to “stabilize groups working on ethics research and products after months of chaos.” Croak “will be the Lead for the Responsible AI Research and Engineering Center of Expertise, she said in a YouTube video announcing her appointment.”
DeepMind Researchers Say AI Poses A Threat To People Who Identify As Queer
Summarizing the findings of a recent DeepMind study “that looked at the positive and negative effects of AI on people who identify as lesbian, gay, bisexual, transgender, or asexual,” VentureBeat (2/18, Johnson) reports that “the impact of AI on people who identify as queer is an underexplored area that ethicists and researchers need to consider, along with including more queer voices in their work.” The DeepMind paper “considers a number of ways AI can be used to target queer people or impact them negatively in areas like free speech, privacy, and online abuse,” as well as in healthcare.
AI Agents Play 'Hide the Toilet Plunger' to Learn Deep Concepts About Life
IEEE Spectrum
Eliza Strickland
February 11, 2021
Researchers at the Allen Institute for AI (AI2) demonstrated that artificial intelligence agents learned the concept of object permanence—that objects hidden from view are still there—by playing hide and seek. The agents, playing as both hiders and seekers, learned the game "Cache" via reinforcement learning. The agents began learning about the environment by taking random actions, like pulling on drawers, and dropping objects in random places. Their game play improved as they learned from outcomes, with the hider, for instance, learning that it had selected a good hiding place when the seeker failed to find the object. Subsequent testing showed that the agents understood the principles of containment and object permanence and were able to rank images based on how much free space they contained. The agents performed as well or better than models trained on the gold-standard ImageNet.
AI Can Use the Veins on Your Hand Like Fingerprints to Identify You
New Scientist
Matthew Sparkes
February 12, 2021
Researchers at the University of New South Wales in Australia developed a technique to identify people using the unique pattern of veins on the back of their hands. They used 500 photos of the hands of 35 people to train a neural network to connect the pattern of veins to a particular subject. The model identified the test subjects with an accuracy rate of 99.8%, then identified four new subjects not included in the original dataset with a 96% accuracy rate. Researcher Syed Shah said vein detection is reliable for people of all ethnicities and is less vulnerable to attacks than existing biometric tests using fingerprints or face recognition. Shah said the technique potentially could be adapted for use with smartphones and CCTV cameras.
*May Require Paid Registration
FLeet: Putting Machine Learning in Your Pocket
EPFL (Switzerland)
Tanya Petersen
February 11, 2021
Researchers from the Swiss Federal Institute of Technology Lausanne (EPFL) and the French National Institute for Research in Digital Science and Technology (INRIA) have demonstrated that mobile devices can perform machine learning in real time as part of a distributed network without compromising device functionality or sharing data. The researchers created FLeet, which combines the privacy of standard the Federated Learning model with I-Prof, a lightweight profiler that predicts and controls the impact of learning tasks on mobile devices, and AdaSGD, an adaptive learning algorithm that is resilient to delayed updates. EPFL's Anne-Marie Kermarrec said, "What we have shown is that if we put all our phones together they start constituting big computing power to match the likes of Google and that gives people alternatives to relying on centralized, powerful computer farms."
Deepfake Detectors Can Be Defeated, Computer Scientists Show for the First Time
UC San Diego Jacobs School of Engineering
February 8, 2021
Computer scientists at the University of California, San Diego (UCSD) demonstrated for the first time that detectors programmed to spot deepfake videos can be beaten. Presenting at the Winter Conference on Applications of Computer Vision 2021 in January, the researchers explained how they inserted adversarial examples into every video frame, inducing errors in artificial intelligence systems. The method also works after videos are compressed, because the attack algorithm estimates across a set of input transformations how the model ranks images as real or fake, then uses this calculation to alter images so the adversarial image remains effective after compression and decompression. The USCD researchers said, "We show that the current state-of-the-art methods for deepfake detection can be easily bypassed if the adversary has complete or even partial knowledge of the detector."
New Study Assesses 40 Years Of Facial Recognition Research
Mashable (2/6, Kraus) reports researchers Deborah Raji of Mozilla and Genevieve Fried of AI Now published Friday the results of their new study assessing 40 years of facial recognition datasets. The MIT Technology Review described “it as ‘the largest ever study of facial-recognition data’ that ‘shows how much the rise of deep learning has fueled a loss of privacy.’” The article examines nine major conclusions from the paper.
AI-Driven Lie Detector Used At EU Borders Faces Scrutiny
Reuters (2/5, Bacchi, Foundation) reported, “A lie detector driven by artificial intelligence and trialled at European Union borders is the focus of a lawsuit that hopes to bring more transparency over the bloc’s funding of ‘ethically questionable’ technology, its proponent said.” European lawmaker Patrick Breyer “is requesting the release of EU Research Agency (REA) documents evaluating the 4.5 million euro ($5.4 million) trial of the use of artificial intelligence (AI) lie detectors to ramp up EU border security.”
Clearview AI's Facial Recognition App Called Illegal in Canada
The New York Times
Kashmir Hill
February 3, 2021
Canadian authorities declared the Clearview AI facial recognition application illegal, with Canada's privacy commissioner Daniel Therrien calling it a tool for mass surveillance. App developer Clearview said it used more than 3 billion photos from social media networks and other public websites to build Clearview AI, currently used by more than 2,400 U.S. law enforcement agencies. Canada's privacy laws mandate obtaining Canadians' consent to use personal data; Clearview claimed it did not require consent to use facial biometric information taken from publicly available photos online. The commissioners balked at the images being used in a manner that the photos' posters had not intended, in a way that could "create the risk of significant harm to those individuals."
*May Require Paid Registration
Chemistry, Computer Science Join Forces to Apply AI to Chemical Reactions
Princeton University
Wendy Plump
February 4, 2021
Chemists and computer scientists at Princeton University collaborated on the development of open-source machine learning software that uses artificial intelligence to optimize chemical synthesis reactions. The software adapts key principles of Bayesian optimization to enable faster, more efficient chemical synthesis. Said lead author Benjamin Shields, who created the Python package, "In designing the software, I tried to include ways for people to kind of inject what they know about a reaction. No matter how you use this or machine learning in general, there's always going to be a case where human expertise is valuable."
Algorithm May Be the Key to Timely, Inexpensive Cyber Defense
Penn State News
Matt Swayne
February 3, 2021
A team of researchers led by The Pennsylvania State University (Penn State) has developed an adaptive cyber defense against zero-day attacks using machine learning. The new technique offers a powerful, cost-effective alternative to the moving target defense method used to detect and respond to cyberattacks. Reinforcement learning enables the decision maker to learn to make the right choices by choosing actions that maximize rewards. Said Penn State's Peng Liu, "The decision maker learns optimal policies or actions through continuous interactions with an underlying environment, which is partially unknown. So, reinforcement learning is particularly well-suited to defend against zero-day attacks when critical information—the targets of the attacks and the locations of the vulnerabilities—is not available."
Supercomputer in Your Bedroom: Researchers Unleash Potential of Desktop PCs to Run Simulations of Mammals' Brains
University of Sussex (U.K.)
Neil Vowles
February 2, 2021
Researchers at the U.K.'s University of Sussex used the latest graphical processing units (GPUs) to give a single desktop PC the capability to perform a large-scale brain simulation that typically requires a supercomputer. The simulations run on the desktop PC consume 10 times less energy than a supercomputer. Simulations using the researchers' GPU-accelerated spiking neural network simulator took up to 35% less time than a previous supercomputer simulation. Said the university's Thomas Nowotny, "This research is a game-changer for computational neuroscience and [artificial intelligence] researchers who can now simulate brain circuits on their local workstations, but it also allows people outside academia to turn their gaming PC into a supercomputer and run large neural networks."
AI Finds More Than 1,200 Gravitational Lensing Candidates
Lawrence Berkeley National Laboratory
Glenn Roberts Jr.
February 2, 2021
A research team including physicists at the U.S. Department of Energy's Lawrence Berkeley National Laboratory (Berkeley Lab) used artificial intelligence (AI) to identify over 1,200 gravitational lensing candidates. The researchers used a sample set of 632 observed lenses and lens candidates, and 21,000 non-lenses, to train a deep residual neural network; the dataset came from the Dark Energy Camera Legacy Survey and Dark Energy Survey. The candidate lenses discovered by the AI can yield insights on the role of dark matter in large celestial objects, and Berkeley Lab's findings, if verified, could more than double the number of known lenses. Berkeley Lab's David Schlegel said, "I really thought it would be many years before anyone would find this many gravitational lenses. It's just amazing to know that you're seeing, very clearly, space itself being warped by a massive object."
'Audeo' Teaches AI to Play the Piano
UW News
Sarah McQuate
February 4, 2021
University of Washington (UW) researchers have developed a system that generates audio from silent piano performances. UW's Eli Shlizerman explained the goal of the research “was to see if artificial intelligence could generate music that was played by a pianist in a video recording." Audeo decodes what transpires in the video and translates it into music, by first detecting which keys are pressed in each frame to produce a diagram; it then converts the diagram into something a music synthesizer would recognize as a sound a piano would make. This step de-noises the data and adds more information, like how strongly each key is pressed and for how long. When the UW team tested Audeo's compositions with music-recognition applications like SoundHound, the apps correctly identified the piece with about 86% accuracy. Said Shlizerman, “We hope that our study enables novel ways to interact with music.”