This week’s list of data news highlights covers April 27-May 3, 2019, and includes articles about an AI system that can mimic Van Gogh and an AI system that maps Saturn’s storms.
1. Predicting Tuberculosis Resistance
Researchers from Harvard University have developed an AI system that predicts a tuberculosis (TB) strain’s resistance to the ten drugs most commonly used to treat it. The system, which the researchers trained on data about 3,600 TB strains resistant to drugs, can predict a TB strain’s resistance to drugs in less than a second with greater than 90 percent accuracy. This system can improve upon current methods to detect TB sensitivity to medication, including culture-based testing that occurs in a lab, which can take up to six weeks because TB grows slowly in lab settings.
2. Using AI to Decide Which Food Imports to Inspect
The U.S. Food and Drug Administration (FDA) is launching a pilot to use AI to decide which food shipments humans should inspect. The system will use a machine-learning model to analyze factors, including the type and supplier of the food, affecting the likelihood a food shipment is hazardous. The United States imported 14 million different food and animal-feed products in 2018, and the FDA will test the AI system against historical results to determine how accurately it flags hazardous shipments.
3. Combining Bach and the Beatles
OpenAI, a nonprofit AI research company, has developed a tool called MuseNet that uses machine learning to generate musical compositions. OpenAI trained Musenet on thousands of pieces of music, allowing MuseNet to discover patterns in harmony, rhythm, and style and predict the upcoming note in a piece given a set of preceding notes. MuseNet can combine styles, such as Bach and The Beatles, and can create music with ten different instruments.
4. Making Phones Weather Sensors
ClimaCell, a weather technology startup based in Boston, has developed a system to use the signals from wireless devices, such as cellphones and IoT devices, to provide forecasts with as much as 60 percent more accuracy than traditional methods. Climacell’s system analyzes the quality of signals from millions of devices as a proxy for weather conditions, including precipitation and air quality, and uses images from street cameras to increase the accuracy of its system.
5. Improving the Performance of Supercomputers
Researchers from Virginia Tech have developed an AI model that helps supercomputers properly balance data processing tasks across thousands of servers. Unbalanced loads can degrade the performance of supercomputers, and the researchers’ system uses a time-series model, which predicts future values based on previously observed values, to predict future application requests with 99 percent accuracy. As a result, the system knows when a task is too great for one server, helping balance the load across the servers that make up the supercomputer.
6. Detecting Cracks in Train Wheels
BNSF Railway, which operates more than 8,000 locomotives and 30,000 miles of track in the United States, is using AI identify to cracked train wheels. The company’s system analyzes 750,000 images of wheels per day taken at seven locations, and the system uses RFID tags to identify trains. Since October, the system has identified 30 cracked wheels that would have been otherwise more difficult to detect.
7. Detecting and Predicting Glaucoma Progression
Researchers from IBM and New York University have developed an AI system that can help detect and predict the progression of glaucoma. Glaucoma can cause visual defects, which is why physicians use visual field tests to map how well patients see, but these tests can be inaccurate because they rely exclusively on patient feedback. The researcher’s AI system estimates a patient’s visual field index from a single image of the optic nerve, allowing a physician to quickly learn a patient’s visual function without multiple time-consuming tests.
8. Creating Van Gogh-Like Paintings
Researchers from the University of Maryland, Adobe, and ByteDance, a Chinese technology company, have developed an AI system called PaintBot that can generate paintings in the style of a particular artist, such as Van Gogh, or in a style itself, such as pointillism. The researchers trained a reinforcement learning algorithm on three to ten paintings per style or artist, and Paintbot can learn to imitate a given painter in roughly six hours. Paintbot learns the color, density, position, size, and order of brushstrokes of a painter or style, allowing it to create new paintings in the desired style based on photos.
9. Predicting Students Who Will Drop Out
Researchers from the Universidad Autónoma de Madrid in Spain and Dimetrical, a data analytics company, have developed a machine learning tool called the Dropout Prevention System to predict the risk of online students dropping out. The researchers trained the system using data from 11,000 students, including data on the age, gender, academic results, fee payment type, and the time and duration of online activity. The system revealed that more activity at night was associated with lower risk scores while data on the length and amount of message interactions between students and teachers was not predictive of a student’s risk of dropping out.
Researchers from the University of Arizona and University College London have developed a deep learning algorithm called PlanetNet that identifies and maps Saturn’s storms. The researchers provided PlanetNet a dataset of infrared data from multiple of Saturn’s storms in 2008, and the algorithm analyzed the data for signs of clustering in the atmosphere’s cloud structure and gas composition. PlanetNet then produced a map that illustrated profound differences between the center of storms and surrounding areas, revealing that the clouds were part of a large upwelling of ammonia ice clouds that surrounded a central storm.
Image: Eric Koch / Anefo