This week’s list of data news highlights covers May 18-24, 2019, and includes articles about an AI system predicting severe respiratory failure and a facial recognition system for pandas.
Sens. Martin Heinrich (D-NM), Rob Portman (R-OH), and Brian Schatz (D-Hawaii) have proposed the Artificial Intelligence Initiative Act to expand the adoption of AI in the United States, create an AI-ready workforce, and foster the responsible use of AI globally. The bill would allocate $2.2 billion of U.S. federal money to developing AI over the next five years, including an extra $1.5 billion to the U.S. Department of Energy to expand its AI research efforts and $100 million a year to the National Science Foundation to create five new centers that promote AI research and education.
2. Predicting Severe Respiratory Failure
Montefiore Health System, a hospital system in New York, has developed an AI system that predicts if patients will develop respiratory failure that requires mechanical ventilation. Montefiore trained the system on the electronic medical records of nearly 70,000 admissions from four hospitals. The AI system analyzes approximately 40 data points per patient, including blood pressure, to make predictions with 64 percent accuracy, which is higher than the 16 percent to 28 percent accuracy of traditional systems.
3. Using Autonomous Vehicles to Help Deliver the Mail
The U.S. Postal Service (USPS) has started a pilot to test if autonomous vehicles can help deliver mail. The trucks, which are built by San Diego-startup TuSimple, are 18-wheel trucks and will make round trips between cities’ distribution hubs, traveling on three freeways and through three states in the process. The vehicles have humans aboard, including one who will drive on any non-freeways.
4. Facial Recognition for Pandas
Researchers from the Chengdu Research Base of Giant Panda Breeding in China, a nonprofit research and breeding facility of giant pandas, Sichuan Normal University in China, and the Singapore Nanyang Technological University have created a facial recognition app that identifies pandas. The researchers trained the app on 120,000 images and 10,000 video clips of pandas, and the app analyzes elements such as the shape of the mouth, size of the ears, and the markings around the eyes to identify individuals. The app can help track the population and activity of panda populations in captivity and the wild.
5. Helping AI Systems Generalize
Researchers from North Carolina State University and Salesforce have developed a framework that helps AI systems better learn new tasks while retaining more of what they have learned from previous tasks. AI systems are subject to catastrophic forgetting, which is when AI systems perform worse at old tasks after researchers train them for new tasks. The researchers’ framework addresses this issue by allowing neural networks to do one of four things when modifying its layers while learning a new task: skip the layer; use the layer in the same way that previous tasks used it; slightly modify the layer; or create a new layer.
6. Creating an AI Robot That Can Work With Humans
Researchers from MIT have developed a robot that uses AI to monitor the muscles signals in a person’s arm to help lift and move objects. The person coordinating with the robot wears small sensors, and a neural network classifies biceps and triceps activity to detect up or down motions. The robot can also detect arm stiffness to hold an object taut.
Samsung has developed an AI system that can generate a lifelike video of a person’s head moving from as little as a single photo or painting of that individual. The researchers trained their system on VoxCeleb, a large dataset of YouTube videos, helping the system learn how human faces move and become incredibly efficient at identifying facial landmarks. This frontloading of the training allows the system to generate clips of new faces with little data.
8. Combining Autonomous Vehicles and AI Robots to Make Deliveries
Ford and Agility Robotics, an Oregon-based startup, have developed a system that combines autonomous vehicles and a bipedal robot to deliver packages. When the autonomous vehicle reaches the delivery address, the robot unfolds out of the trunk of the vehicle to deliver packages that weigh up to 40 pounds. The companies will trial the robot, which uses its own sensors such as LIDAR and detailed mapping data from the autonomous vehicle to navigate, in 2020.
9. Identifying Infection Subtypes
Researchers from the University of Pittsburgh used a machine learning algorithm to find that there are four subtypes of sepsis, a blood infection that kills nearly 270,000 patients in the United States every year. The algorithm analyzed 29 variables, including a patient’s white blood cell count, from the electronic health records of 20,000 patients with sepsis. The algorithm found that in-hospital death rates varied from 2 percent to 32 percent depending on the subtype of sepsis, and this research could foster future work that identifies different therapies for specific subtypes of sepsis.
10. Predicting Adversarial Attacks
Researchers from MIT have developed a method to reduce the effectiveness of adversarial AI attacks, which are attempts to fool AI systems by introducing malicious inputs, such as by using stickers to trick an autonomous vehicle into thinking a stop sign is a 45-mile-per-hour sign. AI systems can identify patterns that correlate with the meaning of the data, such as how the stripes on a large, four-legged animal may correlate with the animal being a tiger, and patterns that form misleading correlations. The researchers reduced their model’s vulnerability to adversarial attacks from 95 percent to 50 percent by choosing which patterns to train their model on, instead of allowing the model to choose.
