10 Bits: the Data News Hotlist
This week’s list of data news highlights covers June 10-16, 2017, and includes articles about a consumer genomics service for babies and an AI system that has mastered Ms. Pac-Man.
Facebook has detailed the many ways it has begun to use AI to identify, analyze, and reduce extremist activity on its platform. Facebook has started experimenting with using AI to analyze the text from accounts removed for praising extremist groups to develop methods for automatically identifying language used to advocate for terrorism. Additionally, Facebook is using learning algorithms that can identify and remove entire clusters of terrorist profiles, and it is continuously improving its algorithms that detect and remove the repeated creation of fake accounts from users already banned from Facebook for extremism.
Researchers at Clemson University and Dartmouth College have developed a prototype headband called Auracle that uses a microphone to monitor when a wearer is eating. Auracle relies on a microphone placed on a user’s skin that can differentiate between the sounds users make when they chew and when they talk or cough with 90 percent accuracy. Auracle could eventually help researchers gather useful data that they can use to study eating disorders and people’s dietary habits.
Two groups of researchers at the Massachusetts Institute of Technology (MIT) have developed a method for using specialized sensors called GelSight to allow robotic arms to mimic aspects of the sensation of touch. GelSight sensors use physical contact to generate a detailed 3D model of an object’s surface. One group of researchers applied GelSight sensors to a robotic arm’s gripper and use machine learning to analyze this data to estimate the hardness of objects it touches. The other group used a similar approach to allow a robotic arm to manipulate small objects with a high degree of dexterity.
A team of researchers from DeepMind and nonprofit AI research firm OpenAI have developed a method for training AI systems that involves humans specifying desirable behavior for a system and only providing occasional feedback, similar to how humans teach toddlers how to perform new tasks. The researchers had humans watch two short clips of an AI system playing simple video games and provide feedback about which clip exhibited better performance, and then fed this data to the AI system to teach it to predict how it should modify its behavior to maximize its progress. Unlike reinforcement learning, which involved a clearly defined reward system, this approach allows an AI system to figure out desirable behavior from humans reviewing just 0.1 percent of its behavior and nudging it in the right direction.
Boston-based Veritas Genetics has begun offering a DNA-sequencing service called myBabyGenome in China to allow parents to learn about their children’s genetic information. myBabyGenome sequences a baby’s entire genome and generates risk scores for 950 serious diseases, provides information on 200 genes that determine how a baby will react to certain drugs, and predicts over 100 physical traits a baby is likely to develop.
Microsoft has developed an AI system that can score the maximum number of points possible on Ms. Pac-Man for the Atari 2600 arcade console, beating all previous human and computer-player records. Ms. Pac-Man is difficult for AI and humans alike to play at high levels due to its lack of predictability, with the highest human score reaching 266,330 points. Using a combination of reinforcement learning and another machine learning technique called divide-and-conquer, Microsoft’s AI was able to achieve the highest possible score of 999,900.
The U.S. Department of Energy has announced $258 million in new funding over three years to expanding its supercomputing capacity and be more competitive with China, which took the top spot for the most powerful supercomputer in the world in June 2016. The funding is part of the United States’ plan to develop an exascale computing system, capable of performing one quintillion calculations per second, by 2021. China and the United States each have 171 systems on the Top 500 list, which ranks the 500 most powerful supercomputers in the world.
Researchers at Vanderbilt University Medical Center have developed a machine learning system capable of predicting whether or not a patient will attempt suicide within two years with between 80 and 90 percent accuracy, and predicting if someone will attempt suicide in one week with 92 percent accuracy. The researchers trained their system on 5,167 case records of patients admitted to the hospital for self-harm or suicidal ideation, and then tested it on cases of patients with no history of suicide attempts in their medical records. Their algorithm also revealed some interesting findings, such as how taking melatonin supplements had a substantial correlation with elevated suicide risk.
Australian startup Lingmo International has developed an earpiece called Translate One2One that uses IBM’s Watson to translate audio between eight different languages in near-real time without needing a data connection. If both participants in a conversation are wearing Translate One2One, it can translate a conversation between Brazilian Portuguese, Chinese, English, French, German, Italian, Japanese, and Spanish.
Members of BetaCity, a volunteer meetup group of citizen scientists, are deploying a network of environmental sensors in Edmonton, Canada that take advantage of the city’s new public low-power, long range (LoRa) network. BetaCity is focusing on measuring air quality initially because Alberta’s environmental agency has a limited number of monitoring stations in the city. All members of the public are allowed to connect their own environmental sensors to the network and view the environmental data the network collects.