This week’s list of data news highlights covers September 12, 2020 – September 18, 2020 and includes articles about using machine learning to eliminate packaging waste and using genomics to reconstruct early transmissions of COVID-19.
IBM and non-profit organization ProMare, which focuses on marine research and exploration, have developed an autonomous ship that collects data from the sea to study global warming, ocean pollution, and marine mammal conservation. To steer autonomously, the ship uses cameras to capture images of ocean hazards, computer vision to identify these hazards, and automation software to follow international collision regulations. An AI model then makes decisions about the ship’s movements based on an analysis of this data.
Researchers from the Moscow Institute of Physics and Technology, Ivannikov Institute for System Programming in Russia, and Harvard Medical School have developed a machine learning model that identifies which retinal tissues grown in vitro can be used in transplants. Typically, engineers can only spot fully developed retinal tissues by inserting a gene into the DNA of cells that causes them to create a unique signal, however this process makes them unsuitable for transplant. The researchers’ method addresses this by using neural networks, algorithms that mimic the way neurons work in the human brain, to identify tissues from conventional light microscope photographs. Their method accurately identified retinal cells 84 percent of the time, whereas humans only accurately identified cells 67 percent of the time.
Facebook has made available automatic captions in 16 languages for Instagram TV, a feature for sharing long-form video on Instagram, using machine learning. Engineers at Facebook trained machine learning models to predict the characters in words using the same technique they used to train previous models that recognize and respond to spoken words, simplifying the models’ training processes. To further prime the system, engineers used public Facebook posts to train the model to adapt to new words like “COVID.”
Amazon has built a machine learning algorithm to eliminate the packaging waste that customer shipments create. The algorithm uses data from product descriptions and customer feedback to identify a product’s category, such as homeware or toys, to decide whether a product can be shipped using a mailer or polybag in place of a cardboard box. Since the algorithm was implemented in 2019, Amazon has reduced the number of boxes it uses from 69 percent to 42 percent, and it has also eliminated 915,000 tons of excess packaging material.
Researchers from Stanford University, Massachusetts Institute of Technology, and the University of Pennsylvania have created a method that predicts the failure rates of automated systems that make critical safety decisions, such as self-driving cars or robots that perform surgery, while simultaneously protecting a company’s trade secrets. The researchers’ method computes the probability of a black box AI system causing dangerous events using a probabilistic sampling approach. With these probabilities, researchers can determine how often dangerous events will occur and how likely an automated system will fail.
The U.S. Library of Congress has developed an AI tool that allows users to search through a dataset of 1.56 million historical newspaper images dating from 1789 to 1963, based on a keyword, date range, or the state in which the newspaper was published. Once an image is returned, the AI system displays related images that would not be detected by conventional search engines. Developers trained the AI tool to sort through 16 million newspaper pages and detect photographs, illustrations, maps, headlines, and advertisements. The tool is now available to the public and can assist archivists investigating historical photojournalism.
Researchers from Harvard Medical School, Georgia Tech, Boston Medical Center, and Massachusetts General Hospital have developed a machine learning tool that detects potential COVID-19 outbreaks. The system analyzes reports of COVID-19 cases and deaths, positive test rates, social distancing policies, face mask regulations, and changes in testing, to predict where an outbreak may occur. The system also uses information from the Centers for Disease Control and Prevention’s Social Vulnerability Index, which measures the preparation level of communities for hazardous events. In addition to detecting potential outbreaks, the system can estimate how many days it takes for cases to double in an outbreak location.
Researchers at the University of Washington in Seattle have reconstructed how people were transmitting coronavirus in the early days of the pandemic using genomics. The team analyzed 453 genetic sequences of coronavirus genomes from the Washington State outbreak between January 19 and March 15 and discovered that a single case of the virus between January 22 and February 10 caused most of the state’s earliest cases. The team also found that one strand of the virus was circulating in the Seattle area for three to six weeks before the first case of community spread emerged on February 28.
Scientists at Salo Sciences, a California tech startup that develops solutions for climate change, are using AI and satellite imaging to generate a more detailed view of the California’s forests that can be used to help firefighters put out fires in the state. The scientists collected detailed images of forests using over 100 satellites and LiDAR, a type of laser technology that uses light to measure distances, and fed this data to the AI tool. With both sets of data, scientists use the AI system to generate a map of fire hot spots in a forest and predict how fires will move. With the tool, California’s fire agency can create forest management plans to prevent future megafires.
Researchers at Harvard Medical School have developed an AI system that predicts the impact of health interventions on the lifespan of mice using their chronological age and biological age, which is the condition of their physical and mental functions. The researchers developed one AI model that predicted mice’s biological age based on their physical frailty and another model that predicted their lifespan. During testing, the AI system accurately predicted whether life-extending interventions or changes in mice’s diet improved their health or life expectancy. Researchers are hopeful this AI system can be used to develop models that measure the effectiveness of life-extending interventions in humans.
Image: Rita E