This week’s list of data news highlights covers January 16, 2021 – January 22, 2021 and includes articles about diagnosing prostate cancer from urine tests and predicting psychosis with machine learning.
1. Creating AI That Can Best Teach Itself
Researchers at the University of California, Berkeley have developed a new approach called PAIRED that allows AI systems to teach themselves. PAIRED works by pairing two AI systems that are almost identical, but have a slightly different set of strengths. A third AI system designs tasks that are just out of reach for one system to solve but are easy for the other. Through trial and error, the less able system can improve its abilities. When tested, AI systems trained with PAIRED solved 80 percent of puzzle problems that no computer had ever solved before, with about one-third of its success coming strictly from the novel training method.
2. Increasing Diversity of Content Exposure on Social Media
Researchers from Aalto University in Finland have developed a new algorithm that social media companies can use to expose users to more diverse content on their platforms. The algorithm works by identifying users whose shared content reaches the greatest number of people and assigning numerical values to what they share. The values represent a position on an ideological spectrum, such as far left or far right. The algorithm then feeds users who are most likely to help the content propagate across the social media network with diverse content, thereby increasing the diversity of content for all users on the platform.
3. Building an Algorithm so Machines Can Better Understand Human Goals
Researchers from the Massachusetts Institute of Technology have developed an algorithm called Sequential Inverse Plan Search (SIPS) to help machines better infer human goals and plans. SIPS works to learn a human’s goals by observing their behavior and forming provisional incremental plans that it revises as it observes more behavior. When tested, SIPS accurately inferred what a human was ultimately trying to do 75 percent of the time and was up to 150 times faster than existing methods. This type of research could eventually be used to improve a range of assistive technologies and collaborative or caretaking robots.
4. Diagnosing Prostate Cancer from Urine Tests
Researchers from the Korea Institute of Science and Technology have developed a technique that uses AI to diagnose prostate cancer from urine tests. The team first developed a sensor system capable of measuring trace amounts of four selected cancer factors in urine for diagnosing prostate cancer. Then they trained an AI system to identify complex patterns and correlations between the detected signals. When tested, the AI system accurately identified 76 urinary samples with 99 percent accuracy in under 20 minutes.
5. Detecting Hidden Earthquakes with AI
Researchers from Stanford University have developed an AI application called Earthquake Transformer that detects small earthquakes. Small earthquakes are often overlooked because they are harder to detect, but they are important in understanding how earthquakes of different magnitudes interact along a fault line. The researchers trained the AI system to identify small earthquakes from the wavelengths recorded on seismograms using one million hand-labeled seismograms recorded over the past twenty years. When tested on a separate set of seismograms from earthquakes in Japan, the system identified two and a half times more earthquakes than humans did.
6. Detecting Fresh Craters on Mars
Researchers at NASA have developed a machine learning algorithm to detect new craters on Mars from imaging data. The team trained the algorithm to identify these craters using 7,000 low and high-resolution photos captured by the Mars Reconnaissance Orbiter satellite. When the team applied the algorithm to a new set of 112,000 orbiter images, the algorithm discovered 70 new craters previously undetected by scientists.
7. Predicting Psychosis with Machine Learning
Clinicians at the Max-Planck Institute of Psychiatry in Germany have used machine learning to predict psychosis, a condition that 2.6 percent of the global population experiences in their lifetime. The algorithm looks at patterns in clinical and biological data to predict psychosis transitions in patients. When tested on 668 patients, 334 of which were diagnosed with psychosis, the method had an accuracy rate of 86 percent.
8. Tracking Elephant Populations Using AI
Conservationists from the University of Bath in England are using observation satellites and AI to spot and monitor endangered elephants. The team trained an algorithm to identify elephants in areas filled with trees and shrubs using 1,000 satellite images. Although using satellites and AI to monitor endangered species is not new, the researchers’ technique better allows conservationists to monitor animals moving through heterogeneous landscapes and eliminates the risk of disturbing animals during data collection processes that usually take place on manned aircrafts.
Researchers from Colorado State University have designed an AI system that trains dogs to obey oral commands without human assistance. The system works by playing a command such as “lie down” or “sit” once it detects the presence of a dog through a camera. The camera records how the dog reacts to the command and the AI system compares the real-time video to a training dataset of images of dogs complying to the command. If the images match, the AI dispenses a treat through a treat dispenser, and if not, no treat is dispensed.
10. Enhancing Well-Being in Japanese Cities
Two Japanese towns called Aizuwakamatsu and Arao are using AI to enhance the well-being of their residents. Aizuwakamatsu has invested in a data sharing platform where residents can submit data about their vital signs or hospital visits to receive recommendations about what lifestyle adjustments they need to make to stay healthy, such as diet changes or exercise. Arao has invested in an AI-based taxi app that calculates the optimal distance between a user’s departure and destination locations to reduce costs and carbon emissions.
Image credit: Centers for Disease Control
