This week’s list of data news highlights covers August 4 – 10, 2018, and includes articles about a an AI system that can spot potholes and a facial recognition system that will help keep security lines short at the 2020 Olympics.
A division of the Dutch police force called Shared Intelligence is developing a system of inexpensive connected sensors, such as air quality sensors, to monitor for signs of environmental crime to help authorities better respond to perpetrators. To expand the area the police monitor, the team will monitor both sensors managed by the police and sensors that concerned citizens placed on their property. The Shared Intelligence team plans to make data from these sensors available as open data to increase the visibility of environmental crimes
University of Waterloo researchers have developed an AI system that can spot potholes, cracks, and other defects in roads in imagery from vehicle-mounted cameras. The system could allow municipal works agencies to automate road assessment to cut costs and monitor roads more frequently, which could enable more timely improvements. The researchers are exploring how the system could help identify faults in bridges and other infrastructure in drone imagery.
The U.S. Defense Advanced Research Projects Agency (DARPA) has launched an initiative called called Media Forensics to developed AI systems that can detect fake videos, such as those that use AI to convincingly swap a person’s face in a video with another person’s, commonly known as “deepfakes.” Participants in Media Forensics have already developed several AI techniques that can spot fake videos, such as by identifying unnatural-looking blinking, or by noting strange head movements or eye colors.
Online education platform Coursera has developed an AI tool that allows subscriber companies to analyze how well their employees are performing in different classes, how their employees’ skills compare to competitors, and identify useful skills that their employees are lacking. Coursera developed the tool by using machine learning to analyze course content and map 40,000 different skills taught on their platform. The tool benchmarks employee skills on a daily basis so managers can better track their employees’ abilities, such as by identifying highly skilled workers that might be underemployed.
Japanese IT company NEC has developed a facial recognition system to improve security efforts at the 2020 Summer Olympic and Paralympic Games in Tokyo. The system will be able to identify athletes, media, staff, and volunteers, as well as over 300,000 attendees by comparing their faces against photo data paired to an ID card with near-field-communication technology. NEC designed the system to reduce wait times at security checkpoints as long lines in the summer heat poses safety risks.
Pharmacological researchers at the University of Illinois and the Mayo Clinic have developed an analytics tool called the Knowledge Engine (KnowEnG) that can predict how a patient will react to different medications based on their genes. KnowEnG uses algorithms that have been effective in data mining activities outside of the medical field and applies them to genomic data to analyze genetic factors that control gene expression and predict how they will respond to different drugs. The researchers developed the tool as part of the National Institutes of Health’s Big Data to Knowledge initiative, which funds efforts to integrate big data into biomedical research.
Researchers at Swiss Federal Institute of Technology have developed a machine learning algorithm that can accurately reconstruct images transmitted as light over a type of fiber optic cable called multimode fibers up to a kilometer long. Multimode fibers can transmit significantly more data than simpler single-mode fibers, as they have several channels that can transmit data simultaneously, and while they are well suited for light-based signals representing raw data, transmitting images is much more challenging. The algorithm can identify images of handwritten numbers transmitted through multimode fiber with 97.6 percent accuracy for a 0.1 meter-long fiber, and with 90 percent accuracy for a kilometer-long fiber. This approach could eventually help improve endoscopic imaging techniques as well as improve the information carrying capacity of fiber-optic telecommunications networks.
Computer Scientists at Arizona State University have developed an identity verification system called FMCode that analyzes users’ hand gestures to authenticate their identify. FMCode uses machine learning algorithms to analyze hand gestures in video or data from wearable sensors and correctly identify a user between 94 and 97 percent of the time. Hand gestures, unlike passwords, will vary every time a user performs them, so the researchers designed FMCode to be able to differentiate between natural variations in gestures from the same user and fraudulent attempts. FMCode could make it easier to authenticate someone’s identify when using a keyboard is impractical, such as in virtual reality applications, or in environments where minimizing contact with other objects is important such as in a lab or operating room.
AI research nonprofit OpenAI, which has developed bots capable of playing the video game Dota 2 has successfully pitted the bots against professional human players and won two of three matches. OpenAI had already demonstrated that its bots, known as the OpenAI Five, could compete with casual human players, but its recent victory shows that its bots, which can train at a rate of 180 years of game time per day, is improving to the point where it can beat some of the best humans.
Students at Fast.ai, which offers free machine learning courses online, have developed an image classification algorithm that outperforms Google in a benchmark called DAWNBench, which evaluates an image classification deep learning algorithm based on speed per dollar of computing power. Google had held the DAWNBench record for a system trained on multiple machines, but the Fast.ai team was able to beat this record by approximately 40 percent, although it used different hardware.
Image: Alan Stanton.