This week’s list of data news highlights covers October 17, 2020 – October 23, 2020 and includes articles about the first multilingual machine translation system and designing cities using machine learning.
Researchers at Riken’s Center for Computational Science and Kobe University in Japan have used Fugaku, the world’s fastest supercomputer, to simulate how coronavirus spreads among people sitting at a dining table with and without their masks on. Researchers discovered that when an infected person speaks, four times as many virus droplets reach the person seated directly in front of them compared to the person seated diagonally across from them. The simulations also revealed that the diner sitting next to the infected person is exposed to five times as many virus droplets than the diner sitting directly across the table.
Researchers at Facebook have open-sourced the first multilingual machine translation system that can translate between any pair of 100 languages without first translating them to English like most other translation models do. Researchers trained the model by grouping languages based on their linguistic, geographic, and cultural similarities, and connected different groups using bridge languages, which are intermediary languages used for translation between groups that do not share a native language or dialect. For example, Hindi, Bengali, and Tamil are bridge languages for Indo-Aryan languages in South Asia. By evaluating all bridge language combinations, researchers produced a dataset of 7.5 billion parallel sentences corresponding to 2,200 different translation directions, such as French to Chinese, allowing the AI model to better preserve linguistic meanings.
Google has partnered with Roostify, a company in San Francisco that has built a digital lending platform, to launch Lending DocAI, an AI service that automates mortgage and loan applications for borrowers and lenders using machine learning models. Lending DocAI is able to parse through document data and extract key information, such as invoice and tax numbers, employer details, addresses, and social security numbers, to speed up the manual document review process that loan approvers typically do. Lending DocAI can also be applied to tax, income, and asset statements and can help mortgage companies comply with regulatory requirements.
The Ocean Cleanup, a Dutch nonprofit organization that has created autonomous boats to clean up oceanic waste, has partnered with Microsoft to use machine learning to identify the different types of waste the autonomous boats collect. Traditionally, staff would manually label and identify over 30,000 images of debris, however, researchers at Microsoft have made this process more efficient by applying a machine learning model that differentiates between images of plastic waste and other debris like leaves or branches. This allows the Ocean Cleanup to more accurately track the waste found in different water regions, and informs how and where it should deploy the boats next.
Researchers at DeepSense, an ocean data analytics hub at Dalhousie University in Canada, have developed a machine learning method for predicting wind speed and wave height measurements that can help port workers safely navigate and transfer vessels, such as cruise ships, during scheduled maintenance sessions or unexpected sensor failures. The researchers used sensor data collected by smart buoys from the past seven years to create predictive models about wind speeds and wave heights. Using the models, researchers hope to create a dashboard that shows changes in the water due to seasonal variances.
Dentsu Inc., a Japanese international advertising company, has created an AI smartphone app that enables consumers to select high-quality tuna using their camera. App developers trained the AI system using the reports of skilled merchants evaluating cross-sections of tuna tails. Based on the merchants’ grading criteria, the developers created a four-stage quality assessment that the algorithm uses to grade a photo of a tuna’s tail. Currently, a conveyor belt sushi chain restaurant is using the app to determine the quality of yellowfin tuna before purchasing.
Sidewalk Labs, an urban innovation company in New York focused on sustainability, has developed Delve, a tool that uses machine learning to generate urban development plans given a set of design constraints. To do this, app developers applied machine learning to a set of core components found in already established neighborhood developments, such as buildings, open spaces, and energy infrastructure, to train the tool to identify the basic features included in every design plan. The tool then explores millions of design possibilities for a given project, measuring the impact of these designs to help development teams arrive at the one that is right for their design constraints, such as budget, location, and square-footage, into the tool.
Scientists at Sydney University in Australia have developed a sensor that uses a light converter and AI to measure the distortion of starlight that occurs when light travels through the Earth’s atmosphere. Traditionally, astronomers identify planets orbiting distant stars by measuring changes in light that occur when planets obstruct the stars and the light they emit, however this is difficult to do from the ground. Now, astronomers can use the measurements sensors generate to directly identify planets orbiting stars that are distant from Earth. The sensor will soon be installed on the Subaru telescope at Maunakea, Hawaii, one of the largest telescopes in the world that uses infrared and optical light.
Scientists at NASA and the University of Copenhagen in Denmark have identified and counted 1.8 billion individual trees in the Sahara Desert from satellite imaging using a deep-learning algorithm. Individual trees are practically invisible to the human eye in satellite images. To train the algorithm to identify and map individual trees, scientists trained the algorithm to recognize what a tree looks like using thousands of images of different types of trees. This new method can help scientists researching the significance of trees in drylands and inform agroforestry techniques climate change specialists create for arid regions.
Howe and Howe Technologies, a vehicle design and fabrication company, has created the first remote-controlled firefighting vehicle in the United States. The robot is equipped with cameras and can pull objects weighing 8,000 pounds, spray water or foam at 2,500 gallons per minute, and plow through debris, including vehicles. The Los Angeles Fire Department currently uses the robot to prevent firefighter deaths in high-risk situations, such as a fire in a structurally unstable building.