This week’s list of data news highlights covers March 21-27, 2020, and includes articles about using supercomputers to combat the coronavirus and developing autonomous drones that can avoid flying objects while traveling at high speeds.
The White House has created the COVID-19 High Performance Computing Consortium, a partnership between the government and technology firms to make supercomputers freely available to coronavirus researchers. The consortium could help speed the discovery of vaccines and treatments because supercomputers have the processing power to predict the effects of drugs, which requires simulating the structures of molecules and their chemical features. At least 16 supercomputers will be available to researchers.
Researchers from Google have developed a neural network that can help design chips for AI systems. The researchers used reinforcement learning to train the network, rewarding it for improvements in performance, power reduction, and area reduction. The system produced a design for Google’s Tensor Processing Unit, a type of integrated circuit, in less than one day. The chip outperformed designs that took human experts weeks to create.
Researchers from MIT have developed a simulation system to train autonomous vehicles. The researchers built the simulator using data about the trajectories of cars driving in the real world, and the simulator leverages reinforcement learning to reward an autonomous system for learning how to navigate on roads safely. The MIT researchers trained an autonomous system for 6,000 miles in the simulator before successfully deploying it on roads in real life.
SparkBeyond, a startup based in New York, used AI to analyze government data about COVID-19 patients’ movements to identify infection hotspots in Italy. The analysis shows that areas such as water fountains, gas stations, and museums correlate with higher rates of infection.
Researchers from Stanford University and the Cedars-Sinai Medical Center in Los Angeles have developed an AI system that can detect how much blood the heart is pumping. The system analyzes ultrasound video frames, and the researchers trained the system using over 10,000 ultrasound videos. The system achieved roughly 95 percent accuracy.
Researchers from Central South University in China and Yale University have developed an AI system that can diagnose the severity of brain tumors. The researchers trained the system using data from the Cancer Imaging Archive, a database of MRIs. The system differentiated between tumor types, such as slow-growing tumors and glioblastomas, an aggressive form of cancer, with 85 percent accuracy.
Researchers from the University of California, Berkeley, and Google have developed a neural network that helps robots reach their desired location in indoor environments. The researchers trained the network in simulation, teaching it to anticipate and react to human motions. The network helped a robot take efficient paths to its desired location in real-world tests.
Researchers from the University of Zurich have developed a system that uses special cameras and algorithms to reduce an autonomous drone’s reaction time down to a few milliseconds, allowing it to avoid other objects while traveling at high speeds. The system uses event cameras, which feed algorithms only pixels that have changing brightness, allowing for faster computation. The system helped a drone traveling 10 meters per second to avoid a ball thrown at it from 3 meters away more than 90 percent of the time.
Researchers from the Swiss Federal Institute of Technology and Bern University Hospital in Switzerland have developed a platform that uses machine learning to predict critical circulatory failure in patients. The platform used 20 variables, including a patient’s blood pressure, pulse, age, and medications, to predict circulatory failure with 90 percent accuracy during a trial. The platform also identified 82 percent of circulatory-failure events more than two hours in advance.
Researchers from the University of Waterloo in Canada and DarwinAI, a Canadian AI startup, have developed a neural network that can detect COVID-19 in chest x-rays. The researchers trained the system using nearly 6,000 x-rays from almost 3,000 patients suffering from various lung conditions, including COVID-19. The researchers released the network to the public to allow other researchers to improve upon it.
Image: Carlos Jones/ORNL