This week’s list of data news highlights covers June 22-28, 2019, and includes articles about an AI system that can predict if stars are likely to have orbiting planets and a tool that can instantly generate machine learning models for data science novices.
MyoKardia, a U.S. cardiovascular medicine company, has shown that a machine learning algorithm can accurately screen for hypertrophic cardiomyopathy (HCM), a condition associated with increased risks of heart failure, stroke, and sudden death. The researchers used smartwatches to gather data on the blood volume changes and electrical pulses of 85 individuals’ hearts. The algorithm correctly identified 95 percent of the individuals with HCM and 98 percent of the individuals who did not have HCM.
Researchers from Google have developed an AI system that can predict the depth of objects in a video better than previous state-of-the-art methods. The researchers trained the system on 2,000 YouTube videos of people performing the mannequin challenge, in which people freeze to imitate being a mannequin, and the corresponding depth data. This research could help develop AI robots or vehicles that can more easily navigate unfamiliar environments.
Researchers in Japan have developed a prosthetic device that uses a neural network to analyze muscle signals and perform finger movements, such as holding a notebook. Seven people using the device were able to perform 10 different finger motions with 90 percent accuracy after the researchers trained the device on only five motions for each finger.
Researchers from MIT have developed an algorithm that can help robots find obscured objects in cluttered scenes. Robots typically turn the data they receive from their sensors into 3D dot representations of objects. However, sensors can provide robots incorrect data that leads to dots that are in the wrong position or are incorrectly spaced, making it more difficult for robots to identify objects. The researchers’ algorithm prunes outlier dots by comparing the features of objects in a template, such as the size and shape of a bunny’s ears, to its input data to remove outliers.
Starsky Robotics, a San Francisco-based autonomous trucking company, successfully tested a heavy-duty commercial truck for 9.4 miles on the Florida Turnpike using a combination of teleoperation and automated driving. A human in a remote facility used a steering wheel, foot pedals, and buttons to drive the truck on and off the highway while the truck’s automation system drove the other 98 percent of the trip.
Researchers from MIT and Brown University have developed an interactive tool called VDS that can instantly generate machine learning models for users with little data science experience. The tool can generate predictive models for tasks ranging from predicting sales revenue to predicting if someone will develop a disease. For example, a user of the tool could drag and drop a dataset concerning the metabolic rates, ages, and disease occurrence of patients, and the tool can then predict whether future patients will develop diseases based on their metabolic rate and age.
Researchers from the Southwest Research Institute, a nonprofit applied research organization, and several U.S. universities have developed an algorithm that can predict whether stars are likely to have planets orbiting them. Stars and planets are created simultaneously and are made of the same materials, and the researchers trained the algorithm on a catalog of thousands of stars, their chemical compositions, and orbiting planets. The researchers’ algorithm identified 360 stars that have a greater than 90 percent probability of hosting a planet.
The Pentagon has developed a system that can analyze a person’s heartbeat and identify them by their heartbeat signature from up to 200 meters away. The system uses lasers to detect the surface movement caused by a heartbeat and can detect heartbeats through regular clothing, such as shirts, but not thick clothing, such as winter coats. The system can identify individuals with up to 95 percent accuracy and could eventually help doctors monitor the conditions of patients without having to connect them to machines.
The researchers and engineers that created the MLPerf benchmark suite, which measures the performance of machine learning models, hardware, and cloud platforms, have created benchmarks for image recognition, object detection, and translation. The benchmarks will test how well a machine learning system can predict a label for a given image, detect an object in an image, and translate sentences between English and German.
Researchers from have developed a neural network that can build 3D simulations of the structural formation of the universe, which usually takes days, in milliseconds. The researchers trained the network on 8,000 traditional universe simulations to teach it how particles interact. Given the direction and distance particles should be moving as the universe expands, the neural network can create simulations with relatively few errors.
Image: European Space Agency