This week’s list of data news highlights covers June 29-July 5, 2019, and includes articles about an AI system that created a flu vaccine and an AI system that can predict drug approvals.
Researchers from the University of Toronto and Preferred Networks, a Japenese firm that uses deep learning to create robots and tools, have developed an AI system that can make a robot made of tree branches walk. The researchers trained the system using reinforcement learning during simulations, rewarding gaits that result in the farthest movement. Combined branches could then walk once the researchers connected them using servo motors, which are devices that produce motion in response to commands.
Researchers from Flinders University in Australia have developed a flu vaccine using artificial intelligence. The researchers’ system consists of one program that learned how to recognize vaccines that were effective against the flu and a second program that developed a list of potential compounds that would be the most effective. The researchers then synthesized the compounds, and a vaccine the AI system created is going through clinical trials in the United States.
Researchers from Penn State, AccuWeather, and the University of Almeria in Spain have developed an AI system that can predict severe weather. The researchers trained the system using more than 50,000 weather satellite images, in which experts had identified and labeled “comma-shaped” clouds, which are strongly associated with cloud formations that lead to hail, thunderstorms, high winds, and blizzards. The researchers’ system can detect comma-shaped cloud withs 99 percent accuracy and predicted 64 percent of severe weather events.
Researchers from Stanford University have created an algorithm that can make wind turbines more efficient. The researchers fed their optimization algorithm five years of data on wind speed, wind direction, and power generation from six turbines on a wind farm. The algorithm suggested turning several of the turbines 20 degrees off the wind, causing choppy air from upwind turbines to be deflected from downwind turbines. The researchers found that this change reduced the power output of the upwind turbines, but increased the overall power output of the collective turbines by as much as 47 percent for certain wind speeds.
Researchers from U.S. and Indian universities have developed a machine learning algorithm that identifies bat species that are potentially hosting the Nipah virus, which can cause severe disease and death in people and is in the midst of an outbreak in the Indian state of Kerala. The researchers developed the algorithm using the data of 500 bat species and their traits, such as their foraging methods, diet, and migration behavior. The algorithm, which identified known species to carry Nipah with 83 percent accuracy, predicts that at least 11 species of bats could be carriers of Nipah in India, significantly more than the one species of bat known to carry the virus in India.
Researchers from Hong Kong University, Texas State University, and C&R Wise, a technology company based in Hong Kong, developed an AI program to count the number of people participating in Hong Kong’s July 1 pro-democracy protest. The size of the crowds at the demonstration is a contentious issue because the size is often measured as a barometer for the strength of the pro-democracy movement. The researchers trained the AI program to distinguish between humans and other objects, such as umbrellas, and ran the program on seven iPads that they placed on two major footbridges. The AI program estimated that 265,000 people marched, while the police stated that 190,000 people participated at the march’s peak.
Researchers from MIT have developed a machine learning model that can predict drug approvals. The researchers used drug-development and clinical-trial data containing thousands of known outcomes, such as if a regulatory body approved a drug to move into the next phase of a clinical trial, to train and test their model. The researchers’ model has an 80 percent probability of correctly predicting if a regulatory body will approve a drug to enter the next phase of a clinical trial.
Researchers from the University of Tübingen have developed a training method to improve the ability of AI systems to classify noisy images correctly. The researchers first tested the ability of algorithms to classify images that contained conflicting cues, such as a silhouette of a cat with the cracked gray texture of an elephant, finding that the algorithms would misclassify the images based on their texture. The researchers then trained the algorithms on images with conflicting textures, forcing the algorithms to rely more on the shape of objects to make identifications, which increased the ability of the algorithms to classify objects in noisy images.
Researchers from the U.S. Department of Energy’s Lawrence Berkeley National Laboratory have shown that an algorithm can uncover new scientific knowledge by scanning millions of papers. The researchers had the Word2vec algorithm analyze the abstracts of more than three million materials science papers, and the algorithm learned concepts such as the periodic table and the crystal structure of metals by examining the relationships between the 500,000 distinct words in the abstracts. The researchers found that the algorithm could use abstracts from up to a specific year, such as the year 2000, to predict future discoveries of new materials scientists eventually made.
Researchers from the University of North Carolina, Chapel Hill, and the University of Maryland have developed an AI system that can detect a person’s emotions by analyzing how they walk. The researchers trained the system on six different datasets of videos of people walking, and they labeled the person in the video as happy, sad, angry, or neutral. The system analyzes features such as a person’s posture, length of stride, and the distance between the hands and the neck to detect emotions such as happiness and anger with roughly 80 percent accuracy.