This week’s list of data news highlights covers August 10-16, 2019, and includes articles about an AI system that predicts food product recalls and an AI system that tracks drug abuse in real-time.
Researchers from MIT have developed an algorithm that has increased the efficiency of school bus routes in Boston. The algorithm created a route map that was 20 percent more efficient than the map humans created, which allowed the Boston public school system to eliminate 50 buses and drive 1 million fewer miles in 2017. The more efficient routing has also reduced carbon dioxide emissions from the busses by 20,000 pounds per day and saved the district $5 million, which it invested into classroom initiatives.
A group of researchers led by an individual from the University of Washington have developed an AI system that can predict certain kinds of food product recalls by the U.S. Food and Drug Administration (FDA). The researchers trained the system on Amazon reviews of products from 2012-2014, which they labeled as corresponding or not corresponding with a product recall. The researchers also annotated the reviews into multiple categories, including if the review implied a consumer became sick or if the review implied the product tasted foul. The AI system predicted three-fourths of product recalls.
Whooshh Innovations, a Seattle firm that creates fish transportation solutions, has developed a passage portal that uses AI to help transfer fish across dams. When fish swim into the portal, a machine takes 18 images of the fish, which an AI system analyzes to determine its length, girth, and if it is from a hatchery, wild, or is an invasive species. The AI system then sorts the non-invasive species into tubes to cross the dam according to their size and releases invasive species.
Researchers from Bioversity International in Africa, India’s Imayam Institute of Agriculture and Technology, and Texas A&M University have developed a smartphone app that uses AI to detect disease in bananas with 90 percent accuracy. The researchers trained the app’s neural network on a dataset of images of diseased bananas in fields in Africa and India. The app can help farmers spot five major banana diseases before they spread, such as black sigatoka, a disease that can cause banana plants to have up to 50 percent lower yields of fruit.
Pindrop, an information-security firm based in Atlanta, has developed voice analytics technology that has helped Discover Financial Services decrease losses from fraud by 10 percent. Pindrop’s technology analyzes 1,380 audio features that indicate the likely origin and transmission of calls, such as over the internet or on a mobile phone. If the system discovers fraud, Discover employees ask callers for a code sent to a device owned by the actual customer.
Nvidia’s AI platform has trained BERT, one of the world’s most advanced natural language processing models, in a record-breaking 53 minutes. Nvidia’s platform also improved BERT’s ability to make inferences to a record-breaking two milliseconds, whereas 10 milliseconds had previously been considered high performance. Nvidia has released its BERT training code and a sample of the optimized BERT model.
Researchers from the University of Massachusetts, Amherst, and Microsoft have developed an AI system that can automatically identify a roof’s solar power potential from satellite imagery. The researchers trained and tested the system on images of roofs from seven different cities, teaching it to determine the geometry of roofs, which determines how many solar panels can fit on a roof. The system also incorporates solar irradiance data to make solar power estimates, which it does with 91 percent accuracy.
Facebook, Deepmind, the University of Washington, and New York University have developed SuperGLUE, a collection of benchmark tasks that test the capabilities of AI systems to perform natural language processing tasks. The benchmark includes tasks that test a system’s ability to understand reason, cause and effect, and answer yes or no questions. SuperGLUE improves on older benchmarks by moving away from trivia-like questions, such as “does a jellyfish have a brain?,” to requiring in-depth answers, such as “how do jellyfish function without a brain?”
Researchers at the New Jersey Institute of Technology have developed DrugTracker, a machine-learning tool that analyzes data from platforms such as Twitter and Reddit to monitor where drug abuse is occurring. The tool analyzes the platforms for slang terms associated with drugs, such as Molly, which is a slang term for the drug MDMA, and the researchers are developing the tool’s capability to analyze images for evidence of drug use. DrugTracker is already monitoring posts in one New Jersey county, and it flags hundreds of relevant posts per day. Counselors and clinicians can use the tool to decide where and how to best allocate their resources.
Researchers from the Tokyo Institute of Technology, multiple UK universities, and the Natural History Museum London have developed an AI system that proved that butterfly species co-evolve similar wing patterns for mutual benefit, which validates a concept known as mimicry theory. The theory, which is biology’s oldest mathematical model, states that toxic butterflies in the same area will evolve to resemble each other to share the loss of some individuals as predators learn their bad taste. The researchers trained the system using 2,400 photographs of butterflies, teaching it to classify photographs by subspecies and by the similarity between wing patterns and colors, finding that the butterfly species borrow characteristics from each other, which validates mimicry theory.