This week’s list of data news highlights covers September 14-20, 2019, and includes articles about identifying cyberbullies online and a system that can predict the size of wildfires.
Researchers led by an individual from the Center for Research and Technology, Hellas, in Greece have developed a machine learning algorithm that can distinguish cyberbullies from other Twitter users with over 90 percent accuracy. The researchers trained the algorithm on a dataset of tweets concerning multiple hashtags, including hashtags related to pay inequality at the BBC and Gamergate, a controversy concerning sexism in video game culture. The researchers found that bully users organize their attacks around societal issues, such as religion, and have fewer followers than other users.
Researchers from ETH Zurich have developed an AI system that can analyze images to spot signs of dark matter. The researchers trained the system on simulated data of what scientists look for when identifying dark matter, which can slightly bend light rays. The researcher’s system is 30 percent more accurate than human scientists at spotting and labeling signs of dark matter in images.
Uber has developed a smartphone-based system called RideCheck that can automatically detect if a car has stopped suddenly. The system uses the GPS, accelerometer, gyroscope, and other sensors in smartphones to detect irregular activity, such as a long stop or car crash. The system then sends an alert to the driver and rider to ask if they are okay.
Researchers from OpenAI have developed AI agents that learned to use tools after playing 500 million games of hide and seek. The researchers used reinforcement learning to improve the agents as they played—hiders and seekers were penalized and rewarded for losing and winning. Over the course of 500 million games, the agents used tools such as boxes and barricades to build forts to hide and seeking agents used ramps to climb over barricades. The agents also began passing each other tools to build forts faster. The research shows that AI agents can develop complex strategies through trial and error.
Researchers from the Allen Institute for AI have developed an AI-enabled tool that can detect if a dietary supplement is unsafe to take in conjunction with another supplement or drug. The tool analyzes a database of health research papers to find potential concerning interactions between supplements and medications. The tool is free to use, and the researchers are updating the database with additional papers.
IBM has developed a 53-qubit quantum computer that is the most powerful quantum computer openly available for researchers and companies to use via the cloud. IBM already has several 20-qubit quantum computers, and the new computer will allow researchers to run more complex experiments and for larger cloud deployments of applications.
Researchers from Los Alamos National Laboratory in the United States and PSL Research University in Paris have developed a machine learning algorithm that can detect slow slip earthquakes, which are earthquakes that can occur over several weeks. The researchers trained the algorithm on data of slow slip earthquakes occurring between 2007 and 2013, finding that the algorithm could then predict four out of the five slow slips that occurred in the Cascadia forests in the Pacific Northwest of the United States between 2013 and 2018.
Researchers from the University of California, Irvine have developed a machine learning system that can predict how large a wildfire will grow at the time of its ignition. The researchers trained the system on air moisture data from more than 1,000 fire events that occurred between 2001 and 2017 in Alaska. The system predicts 40 percent of ignitions that lead to large fires, outperforming more systems that rely on multiple variables.
Researchers from Cornell University and several South American organizations have developed an AI model that can predict the best locations to build hydropower dams to minimize carbon emissions. There are hundreds of proposals to build dams in the Amazon basin, and the researchers model identifies the combination of dams that would produce the lowest amount of carbon emissions given an energy production target. The model found that choosing dams at higher elevations typically minimizes carbon footprints.
Organizations that have significant amounts of user-generated content are using Amazon’s computer vision service Rekognition to automatically delete or blur parts of unsavory images. For example, the food delivery service Deliveroo uses Rekognition to automatically delete reviews that contain inappropriate content, such as nudity. Rekognition uses deep neural networks that first process top-level information such as an object’s shape before analyzing increasingly specific details to classify images.
Image: John McColgan