This week’s list of data news highlights covers July 11-July 17, 2020, and includes articles about using court records to find disparities in the judicial system and an AI-enabled system that can monitor the health of ecosystems.
The Mount Sinai Health System in New York has developed several machine learning algorithms to help doctors make assessments about COVID-19 patients. The algorithms use data from patients’ electronic medical records to generate predictive scores, such as if a patient is likely to need a ventilator. The scores become part of a patient’s medical records, and doctors can use the scores to identify critical issues and optimize hospital resources.
Researchers from Northwestern University have used an algorithm to find disparities in how often judges grant waivers to the $400 fee to file a lawsuit. The algorithm analyzed court records, finding that the waiver rate could vary from 20 to 80 percent, depending on the judge.
OpenAI, a San Francisco-based research lab, has created an AI-enabled system that can generate a complete image using only half an image’s pixels. The system uses GPT-2, an AI model that can generate essays using only a single sentence as input. OpenAI designed GPT-2 to analyze one-dimensional data, such as a string of words. To develop the new model (iGPT), the organization represented images as single sequences of pixels, finding the model could predict an image’s second half of pixels realistically.
Visa is using an AI-based system to prevent $25 billion in annual fraud. The system uses neural networks to analyze 500 risk factors to provide all transactions a fraud-likelihood score. The system analyzed more than 100 billion transactions last year.
Firefighters in Utah are using an AI system to detect fires early on in their development. The system uses thermal imaging cameras and real-time image processing to detect fires. The system helped firefighters identify and extinguish an early stage brushfire in early July.
Researchers from the Imperial College London, University of Sydney, and Cornell University have developed a neural network that can detect abnormal sounds in ecosystems, which can help researchers identify illegal activities. The researchers trained the network to learn the frequency and structure of sounds, such as birds’ calls, using unsupervised learning. This approach allowed the system to learn sounds that are typical of an ecosystem without requiring the researchers to teach it sounds that are atypical, such as logging.
BreathResearch, a startup based in California, is developing a system that helps doctors track the progression of respiratory diseases by analyzing a patient’s breathing. The firm is developing the system using data from 100 healthy and diseased subjects in Mayo Clinic and the University of Wisconsin-La Crosse studies. This process helped BreathResearch build a library of sound data related to specific disease types and severity.
FamilyMart, a Japanese convenience store chain, will begin using a robot to stack shelves in August. Humans will operate the robot until its AI-enabled system learns the movements it must perform. Telexistence, the firm that developed the robot, trains it to mimic human movements using virtual reality headsets and motion-sensor controls.
Researchers from Ubisoft, a video game company based in France, Google, and Macquarie University have developed a machine learning system that can translate hieroglyphs. The researchers trained the system on 80,000 drawings of hieroglyphs. Google has incorporated the system into an app called Fabricius to foster the study of ancient languages.
Researchers from the University of Illinois at Urbana-Champaign have developed a method to train deep learning systems to be less susceptible to adversarial attacks—attempts to fool AI systems by introducing malicious inputs. The researchers developed the process for systems that reconstruct images, such as systems that create MRI scans from x-ray data. The method involves using one algorithm to generate adversarial examples to fool the reconstruction algorithm to include false inputs in its images. The recreation algorithm then successively tweaks itself until it creates images that accurately reflect ground truth data.