This week’s list of data news highlights covers February 22-28, 2020, and includes articles about detecting endangered whales and an AI system that helps autistic children develop social skills.
Researchers from Columbia University have created a synthetic finger that combines machine learning, LEDs, and photodiodes—semiconductors that convert light into electricity—to provide robots a sense of touch. Reflective silicone overlays the LEDs and photodiodes, and the photodiodes detect changing light levels when the finger touches something. A machine learning algorithm then analyzes 960 signals the LEDs and photodiodes create to “see,” rather than feel, the location and intensity of the touch.
Researchers led by an individual from the U.S. National Oceanic and Atmospheric Administration are using underwater drones equipped with acoustic sensors to detect and map the presence of endangered whales in the North Atlantic. The sensors capture the calls of whales and can provide real-time location data. The researchers’ system also automatically analyzes the audio data to identify the type of whale.
Researchers from MIT have developed a system to test the robustness of natural language processing models, which attempt to understand human language. The system replaces words in a given text with synonyms to test a model’s ability to classify text into different news topics, as real or fake news, and as a positive or negative. The researchers found that the system could reduce the accuracy of some high-performing natural language processing models to below 20 percent by changing less than 20 percent of the original words.
All Nippon Airways, Japan’s largest airline, is using an AI-enabled device called Pockettalk that provides translations in real time. The device, which is smaller than a smartphone, can translate 74 different languages and allows the airline’s staff to better communicate with visitors. The device can understand slang and idiomatic phrases.
Cleveland Clinic is developing an AI model to predict a patient’s probability of readmitting to the hospital. The model uses medical information and data on admission records, such as the length of time between previous admissions, to make its predictions. The hospital has started 20 AI-related research projects since opening its Center for Clinical Artificial Intelligence in 2019, including a project to predict patient responses to chemotherapy.
Microsoft, online gaming platform Roblox, and instant-messaging service Kik have developed an AI system that can detect messaging conversations between predators and minors. The AI system analyzes conversations to assign them a risk level, identifying patterns that can indicate inappropriate discussions, such as one party asking the other to move to another platform for direct messaging. If the system identifies a conversation as high risk, it notifies a human moderator.
Indian startup SeeHow has developed sensors for cricket balls and bats to measure and improve player performance. The sensors can track the speed, spin, and distance of the ball and the swing speed, angle, and the area a bat makes contact with the ball. This data streams to an app, which allows players to track their progress over time.
Researchers from the University of Southern California have developed an AI system that can help autistic children develop social skills. The system combines a robot and a machine learning model that analyzes audio and video data, including eye contact and dialogue, to detect if a child is engaged in an activity. The researchers placed the robot in children’s homes for a month, finding that interacting with the robot taught the children social skills, including taking turns to talk and looking at someone when talking. The researchers also found that many of the children improved their empathy towards other humans and became more engaged in conversations with their family.
The U.S. Internal Revenue Service is using AI to both increase the number of individuals that pay their taxes and to detect high-income individuals that do not file returns. For example, the agency is using AI to analyze their employees’ notes to discover the combination of formal notices and other ways of contact that are likely to convince a taxpayer who owes money to comply.
Google has developed a new email malware scanner that utilizes deep learning to increase the number of malicious Microsoft Office documents the company detects by ten percent. The scanner processes more than 300 billion attachments each week, identifying common attack patterns. The scanner has excelled at flagging bursts of malicious documents botnets have sent.
Image: Whit Welles