This week’s list of data news highlights covers August 29 – September 4, 2020, and includes articles about developing a tool to detect deepfakes and conversing with Holocaust survivors through AI-created digital images.
To improve food safety of seafood, 94 percent of which comes from abroad, the U.S. Food and Drug Administration (FDA) is using an AI tool to screen seafood shipments. Staff use the screening tool to decide which shipments to inspect and sample for testing. Researchers trained the screening tool using past seafood shipments that FDA staff rejected or further examined for authenticity. With better screening, the FDA can create a more traceable food system to protect public health.
The Chicago Police Department has developed a system that monitors the mental health of its police officers and flags those who are more likely to cause harm to themselves or others before it happens. The system uses departmental data such as complaints about an officer’s controlled substance usage substances, off-duty behavior, and sustained use of excessive force. Once the system flags an officer’s record, supervisors will have the option to connect them with counseling and therapeutic services.
The U.S. Tennis Association is using two AI platforms to enhance the fan experience at its annual tennis tournament, the U.S. Open. One platform called Open Questions uses natural language processing to curate debates on hot tennis topics. Fans can ask debatable questions, such as whether a specific player is the greatest of all time, and the platform will analyze more than 40 million pieces of data from sports and analytical articles to present the information in a pros versus cons argument format. The second AI platform generates factsheets for fans before matches by collating performance statistics on the competing players and summarizing this information in a readable format.
Microsoft has launched a detection tool to spot deepfakes, which are realistic-looking images and videos that have been manipulated to portray someone doing or saying something that never happened. The tool uses AI to detect signs of manipulation that go undetected by the human eye, such as grayscale pixelation and image fading, and produces a confidence score describing the likelihood the media is a deepfake. Microsoft trained the tool using a public dataset of 1,000 deepfaked videos and tested the tool’s machine learning models against Facebook’s even larger database of artificially created face images.
Google has expanded its AI flood warning system to cover all of India and parts of Bangladesh. The system uses data about river levels, rainfall patterns, and flood simulations to create new prediction models for when, where, and how much flooding might occur in these regions. Researchers say that in 90 percent of cases, the models created will accurately predict the level of flood water within 15 centimeters. With the warning system in place, flood forecasting alerts will now reach 200 million people in India and 40 million people in Bangladesh.
Google has released an anonymized dataset of symptom searches from each U.S. county to help researchers predict future COVID-19 hot spots. The dataset dates back to 2017 and includes 400 symptoms and conditions beyond those commonly associated with the virus, such as stress and diabetes, to explore previously unrecognized impacts of the disease on individuals with underlying health conditions.
The U.S. Army is using a machine learning system from Sparta Science, a movement-based software company, to prevent and predict injuries in soldiers. The system uses sensor-equipped scales to measure a soldier’s core and lower strength based on exercises like planks and jumping jacks. Using this data, the system predicts the risk an individual has for getting certain musculoskeletal injuries. Each system’s diagnostic test only takes five minutes and can also be used to identify weak areas of the body that need to be strengthened to prevent injury.
Researchers from Carnegie Mellon University and the University of Pittsburgh Medical Center in Pennsylvania have developed a machine learning approach that examines placentas delivered after birth for any indication of complications in future pregnancies. Using data from slides of diseased and healthy placentas, the algorithm identifies each blood vessel on the slide and assesses it for characteristics of disease. If the algorithm identifies just one vessel as diseased, the algorithm will flag the entire slide. Researchers are hopeful the algorithm can identify mothers at risk of preeclampsia, a life-threatening condition that causes high blood pressure and reduces organ functions.
Researchers at Columbia University, the Massachusetts Institute of Technology, and IBM have trained a machine learning model to identify and categorize abstract video concepts, such as folding clothes and hammering nails. The researchers trained the model using words and images embedded with information about their interrelationships in order to enhance the model’s ability to group abstract videos based on their similarities. When tested, the model performed better than humans in two scenarios: picking which video clips completed a given abstract set and identifying which videos did not belong in a given abstract set.
The University of Southern California Shoah Foundation, a nonprofit organization dedicated to making audio-visual interviews with Holocaust survivors, is using AI to preserve stories and accounts of the Holocaust. The foundation used 20 cameras to film interviews with 21 survivors, capturing over 2,000 questions and answers to create interactive digital exhibits that are available at select Holocaust Museums in the United States.
Image: Catherine Avalone