Dr. T's AI brief
Daniel Tauritz
Rapid Covid-19 Test the Result of University-Industry Partnership
UC Davis Health
April 26, 2021
Researchers at the University of California, Davis (UC Davis) have developed a rapid Covid-19 test that combines mass spectrometry with robotics and an automated machine learning platform to detect SARS-CoV-2 in nasal swabs. The test, which takes about 20 minutes, was found to be 98.3% accurate for positive tests and 96% for negative tests, matching or outperforming many existing Covid-19 screening tests. The researchers used a mass spectrometer from Shimadzu Scientific Instruments to ionize 226 nasal swabs, and the hundreds of peaks and signals produced by the ionized particles were analyzed by the automated machine learning platform MILO (Machine Intelligence Learning Optimizer) to detect patterns that correspond to the presence or absence of the virus in the samples. SpectraPass, a startup launched by Allegiant Travel Company's Maurice J. Gallagher, Jr., a UC Davis alumnus, will produce the testing system.
AI's Carbon Footprint Is Big, But Easy to Reduce, Google Researchers Say
Fortune
Jeremy Kahn
April 21, 2021
Researchers at the University of California, Berkeley and Google have released the most accurate estimates to date for the carbon footprint of large artificial intelligence (AI) systems. They determined OpenAI's powerful language model GPT-3, for example, produced the equivalent of 552 metric tons of carbon dioxide during its training. The researchers found the carbon footprint of training AI algorithms depends on their design, the computer hardware used to train them, and the nature of electric power generation in the location where the training occurs; changing all three factors could lower that carbon footprint by a factor of up to 1,000. A reduction by a factor of 10 could be achieved through the use of "sparse" neural network algorithms, in which most of the artificial neurons are connected to relatively few other neurons.
AI Tool Tracks Evolution of Covid-19 Conspiracy Theories on Social Media |
3D Motion Tracking System Could Streamline Vision for Autonomous Tech
University of Michigan News
April 23, 2021
A real-time, three-dimensional (3D) motion tracking system developed by University of Michigan (U-M) researchers integrates transparent light detectors with advanced neural network techniques, which could potentially replace LiDAR and cameras in autonomous technologies. Graphene photodetectors absorb only about 10% of the light they receive, enough to produce images that can be reconstructed via computational imaging. Deep learning algorithms enable motion tracking, and the system offers scalability. U-M's Dehui Zhang said the blend of graphene nanodevices and machine learning algorithms "combines computational power efficiency, fast tracking speed, compact hardware, and a lower cost compared with several other solutions."
ML Model Generates Realistic Seismic Waveforms
Los Alamos National Laboratory News
April 22, 2021
The SeismoGen machine learning model can generate high-quality synthetic seismic waveforms, according to researchers at the U.S. Department of Energy's Los Alamos National Laboratory (LANL). The team designed SeismoGen based on a generative adversarial network. Once trained, the SeismoGen model can produce realistic seismic waveforms of multiple labels. The LANL researchers applied the model to actual Earth seismic datasets in Oklahoma. LANL's Youzuo Lin said, "Through a sequence of qualitative and quantitative tests and benchmarks, we saw that our model can generate high-quality synthetic waveforms and improve machine learning-based earthquake detection algorithms."
AI Tool Calculates Materials' Stress and Strain Based on Photos
MIT News
Daniel Ackerman
April 22, 2021
A technique for rapidly assessing material properties like stress and strain, based on an image of the internal structure, has been developed by Massachusetts Institute of Technology (MIT) researchers. The team trained a Generative Adversarial Neural Network with thousands of paired images—respectively depicting a material's internal microstructure subject to mechanical forces, and its color-coded stress and strain values; the network iteratively determined relationships between a material's geometry and its ensuing stresses using principles of game theory. MIT's Markus Buehler said the computer can essentially predict the various forces that act on the material, as opposed to the conventional way, in which "you would need to code the equations and ask the computer to solve partial differential equations. We just go picture to picture." Buehler said the network is well-suited for describing material properties, as it can process data through a series of convolutions, which analyze the images at progressively larger scales.
A Growing Problem of 'Deepfake Geography': How AI Falsifies Satellite Images
University of Washington
Kim Eckart
April 21, 2021
Researchers at the University of Washington (UW), Oregon State University, and Binghamton University used satellite photos of three cities and manipulation of video and audio files to identify new methods of detecting deepfake satellite images. The team used an artificial intelligence framework that can infer the characteristics of satellite images from an urban area, then produce deepfakes by feeding the characteristics of the learned satellite image properties onto a different base map. The researchers combined maps and satellite imagery from Tacoma, WA, Seattle, and Beijing to compare features and generate deepfakes of Tacoma, based on the characteristics of the other cities. UW's Bo Zhao said, "This study aims to encourage more holistic understanding of geographic data and information, so that we can demystify the question of absolute reliability of satellite images or other geospatial data. We also want to develop more future-oriented thinking in order to take countermeasures such as fact-checking when necessary."
Study Explores Deep Neural Networks' Visual Perception
The Hindu (India)
April 21, 2021
A study by researchers at the Indian Institute of Science (IISc) compared the performance of deep neural networks to that of the human brain in terms of visual perception. The team analyzed 13 different perceptual effects, revealing previously unknown qualitative distinctions between deep networks and the brain. IISc's Georgin Jacob observed a surprising local advantage among neural networks over the brain, in that they concentrate on the finer details of an image first. An IISc press release said identifying these differences can lead to better-performing neural networks that are resistant to adversarial attacks.
Gold Digger: Neural Networks at the Nexus of Data Science, Electron Microscopy |
Stanford Researchers Use AI to Empower Environmental Regulators
Stanford News
Rob Jordan
April 19, 2021
Stanford University researchers have demonstrated how artificial intelligence combined with satellite imagery creates a low-cost, scalable method for finding and monitoring otherwise hard-to-oversee industries, which environmental regulators could employ to spot violators. Previous research has tended to focus on wealthy countries, while the Stanford-led work concentrated on Bangladesh, where localization and enforcement of environmental regulations related to highly pollutive brick kilns is difficult. In partnership with the International Center for Diarrheal Disease Research, Bangladesh, the team devised a deep learning algorithm that not only identifies whether satellite images contain such kilns, but also learns to locate kilns within the image. The algorithm can reconstruct kilns fragmented across multiple images, identify multiple kilns within a single image, and differentiate between sanctioned and illegal kiln technologies based on shape classification.
Facebook Releases Dataset To Address Algorithmic Bias
The Wall Street Journal 
(4/8, McCormick, Subscription Publication) reports Facebook has made publicly available a dataset designed to help AI researchers evaluate their computer vision and audio models for potential algorithmic bias. The dataset, called Casual Conversations, has videos of approximately 3,000 participants of various skin tones sharing their age and gender.
AI Developers: Chatbots Will One Day Be Able To Offer Financial Management Services
The Wall Street Journal (4/7, Carpenter, Subscription Publication) reports that according to AI developers, in the future, chatbots will be able to offer banking services and financial management advice. The technology will someday be able to generate personalized advice based on individual questions and situations.
Researchers Believe AI In Education Will See “Significant Growth” Over Next Decade
The Hechinger Report (4/8, Salman) reports that researchers at the Digital Promise-led Center for Integrative Research in Computing and Learning Sciences (CIRCLS) “believe that over the next five to 10 years, AI in the education space will see a significant growth.” In a CIRCLS report, researchers write that they anticipate “AI will come to greatly impact teaching and learning dramatically in the coming years.” They urge educators to begin planning now for “how to best develop and use AI in education in ways that are equitable, ethical, and effective and to mitigate weaknesses, risks, and potential harm.” Jeremy Roschelle, co-principal investigator at CIRCLS, said that focus on AI ed-tech should be on “supports, and tools that assist teachers.” For example, “new tools could include an AI-powered virtual teaching assistants that help teachers to grade homework and provide real-time feedback to students, or that assist teachers in ‘orchestrating and organizing social activity in the classroom,’ Roschelle said.”
Daniel Tauritz
An Artificial Neural Network Joins the Fight Against Receding Glaciers
Columbia University
Daniel Burgess
May 5, 2021
An artificial neural network developed by University of California, Irvine (UCI) researchers can autonomously recognize and quantify the edges of glaciers in satellite images with far greater reliability than humans. The team trained the Calving Front Machine (CALFIN) network on tens of thousands of images; afterwards, CALFIN could measure calving fronts to within an average of 82 meters (269 feet) from their true locations, outperforming previous models, which incurred errors of more than twice that distance on the dataset. Said William Colgan of the Geological Survey of Denmark and Greenland, “I think machine learning now offers a robust way of upscaling a handful of site-specific and process-level observations to tell a bigger regional story."
An Uncrackable Combination of Invisible Ink, AI
American Chemical Society
May 5, 2021
Researchers have printed complexly encoded data using a carbon nanoparticle-based ink that can be read only by an artificial intelligence (AI) model when exposed to ultraviolet (UV) light. The researchers created the ‘invisible’ ink, which appears blue when exposed to UV light, using carbon nanoparticles from citric acid and cysteine. They then trained an AI model to identify symbols written in the ink and illuminated by UV light, and to use a special codebook to decode them. The model, which was tested using a combination of normal red ink and UV fluorescent ink, read the messages with 100% accuracy. The researchers said the algorithms potentially could be used for secure encryption with hundreds of unpredictable symbols because they can detect minute modifications in symbols.
ML Accelerates Cosmological Simulations
Carnegie Mellon University
Jocelyn Duffy
May 4, 2021
Researchers at Carnegie Mellon University (CMU), the Flatiron Institute, the University of California, Riverside, and the University of California, Berkeley, have trained a neural network-based machine learning algorithm to upgrade a cosmological model's resolution dramatically. The trained code renders full-scale, low-resolution models as super-resolution simulations containing up to 512 times as many particles. This approach can produce a high-resolution model of a cosmological region encompassing about 500 million light years and 134 million particles in just 36 minutes on a single processing core, while existing methods would require 560 hours (more than 23 days). CMU's Rupert Croft said, "By incorporating machine learning, the technology is able to catch up with our ideas."
'Brain-Like Device' Mimics Human Learning in Major Computing Breakthrough
The Independent (U.K.)
Anthony Cuthbertson
April 30, 2021
A device modeled after the human brain by researchers at Northwestern University and the University of Hong Kong can learn by association, via synaptic transistors that simultaneously process and store information. The researchers programmed the circuit to associate light with pressure by pulsing a light-emitting diode (LED) lightbulb and then applying pressure with a finger press. The organic electrochemical material enabled the device to construct memories, and after five training cycles it associated light with pressure and could detect pressure from light alone. Northwestern's Jonathan Rivnay said, "Because it is compatible with biological environments, the device can directly interface with living tissue, which is critical for next-generation bioelectronics."
AI, Captain! First Autonomous Ship Prepares for Maiden Voyage
France24
April 30, 2021
The autonomous ship Mayflower 400 is preparing for a transatlantic journey from England to Plymouth, MA. The solar-powered, radar- and camera-equipped trimaran has an onboard artificial intelligence which learned to identify maritime obstacles by analyzing thousands of photos. IBM’s Rosie Lickorish said the unmanned craft provided an advantage in the "unforgiving environment" of the open ocean; "Having a ship without people on board allows scientists to expand the area they can observe." The Mayflower 400 will analyze marine pollution and plastic in the water, as well as tracking aquatic mammals; the data it collects will be released for free.
Breakthrough Army Technology is Game Changer for Deepfake Detection
U.S. Army Research Laboratory
April 29, 2021
Researchers at the U.S. Army Combat Capabilities Development Command's Army Research Laboratory and the University of Southern California (USC) have developed a deepfake detection method for supporting mission-essential tasks. The team said DefakeHop's core innovation is Successive Subspace Learning (SSL), a signal representation and transform theory designed as a neural network architecture. USC's C.-C. Jay Kuo described SSL as "a complete data-driven unsupervised framework [that] offers a brand new tool for image processing and understanding tasks such as face biometrics." Among DefakeHop's purported advantages over current state-of-the-art deepfake video detection methods are mathematical transparency, less complexity, and robustness against adversarial attacks.
Neural Nets Used to Rethink Material Design
Rice University News
April 30, 2021
A technique developed by researchers at Rice University and Lawrence Livermore National Laboratory uses machine learning to predict the evolution of microstructures in materials. The researchers demonstrated that neural networks can train themselves to predict a structure's growth in a particular environment. The researchers trained their neural networks using data from the traditional equation-based approach to predict microstructure changes and tested them on four microstructure types: plane-wave propagation, grain growth, spinodal decomposition, and dendritic crystal growth. The neural networks were 718 times faster for grain growth when powered by graphic processors compared to the prior algorithm, and 87 times faster when run on a standard central processor. Rice's Ming Tang said the new method can "make predictions even when we do not know everything about the material properties in a system," and will be useful in designing more efficient batteries.
ML Algorithm Helps Unravel the Physics Underlying Quantum Systems
University of Bristol News (U.K.)
April 29, 2021
Scientists in the Quantum Engineering Technology Labs (QETLabs) of the U.K.’s University of Bristol have designed an algorithm that provides insights into the underlying physics of quantum systems. The algorithm is an autonomous agent, using machine learning to reverse-engineer Hamiltonian models to overcome multiple complexities; it designs and conducts experiments on a targeted quantum system, with the resulting data fed back to the algorithm. The algorithm then proposes candidate Hamiltonian models to characterize the target system and differentiates them using statistical metrics. QETLabs' Anthony Laing said the researchers "have potentially turned a new page in scientific investigation by bestowing machines with the capability to learn from experiments and discover new physics. The consequences could be far-reaching indeed."
Advanced Core Processing: Robot Technology Appealing for Apple Growers
Monash University (Australia)
April 28, 2021
Researchers at Australia's Monash University have developed autonomous robotic technology capable of harvesting apples. At full capacity, the robot can identify, pick, and deposit an apple in as little as seven seconds, with a median rate of 12.6 seconds per apple. Trials showed the robot could harvest over 85% of reachable apples within a canopy as identified by its vision system, with less than 6% of the harvest damaged by stem removal. Monash's Chao Chen said the vision system uses deep learning to identify apples within its range, and to identify and categorize obstacles like branches. Said Chen, "We also implemented a 'path-planning' algorithm that was able to generate collision-free trajectories for more than 95% of all reachable apples in the canopy."
Poll Finds “AI-Powered” Cyberattacks Expected To Increase
TechRadar (4/12) reports new research has “said AI-powered software will soon be powerful enough spearhead advanced cyberattacks, prompting IT teams to deploy smarter security solutions, themselves.” Polling 300 C-level executives on their views of the future cybersecurity landscape, cybersecurity AI company Darktrace “found that almost all respondents (96%) are preparing for an onslaught of AI-powered cyberattacks.” Nearly two-thirds (68%) are “under the impression that cybercriminals will be deploying AI on impersonation and spear-phishing attacks.” To prepare for future attacks, most of the executives “polled for the report said they started deploying AI-powered defenses, mostly because they don’t believe (60%) humans are a match for automated cyberattacks, even if they could find enough, due to the ever-growing talent drought.” They also “believe current security solutions are a liability because they’re unable to anticipate new attacks.”
Microsoft Looking To Acquire AI Healthcare Record Firm Nuance
Bloomberg (4/12, Bass, Baker, Porter) reports, “Microsoft Corp. is making a massive bet on health-care artificial intelligence” with the purchase of Nuance Communications, “the company tied to the Siri voice technology.” The companies have been working together “for two years on AI software that helps clinicians capture patient discussions and integrate them into electronic health records, and combining the speech technology company’s products into its Teams chat app for telehealth appointments.”
Also reporting are Reuters (4/11, Hu) and the Financial Times (4/11, Fontanella-Khan, Waters, Subscription Publication).
BMW Planning Virtual Factory With AI To Improve Its Assembly Lines
Wired (4/12, Knight) reports that before any of BMW’s EV drivetrains “roll off the production line” at its factory in Regensburg, Germany, “the entire manufacturing process will run in stunningly realistic detail inside a virtual version of the factory.” BMW Production Strategy Lead Markus Grüeneisl said that “the simulation allows managers to plan the production process in greater detail than was previously possible.” It is also part of the automaker’s “plan to use more artificial intelligence in manufacturing.” Grüeneisl added that “machine-learning algorithms can simulate robots performing complex maneuvers to find the most efficient process.”
MIT Researchers Say AWS Attempt To “Discredit” Study About Facial Analysis Employed “Misogynoir” Tactics
In a more than 3,200-word article, VentureBeat (4/10, Johnson) said that after a 2019 research paper “demonstrated that commercially available facial analysis tools fail to work for women with dark skin, AWS executives went on the attack,” and “attempted to discredit study coauthors Joy Buolamwini and Deb Raji in multiple blog posts.” VentureBeat added that “according to the Abuse and Misogynoir Playbook, published earlier this year by a trio of MIT researchers, Amazon’s attempt to smear two Black women AI researchers and discredit their work follows a set of tactics that have been used against Black women for centuries.” According to VentureBeat, the coauthors of the Playbook argue Amazon similarly used “misogynoir” tactics “to disparage former Ethical AI team co-lead Timnit Gebru after Google fired her in late 2020 and stress that it’s a pattern engineers and data scientists need to recognize.”