This week’s list of data news highlights covers June 15-21, 2019, and includes articles about an AI system that can predict a person’s gestures and AI system that can detect deepfakes.
1. Releasing an Updated AI R&D Strategy
The White House Office of Science and Technology has released an updated National Artificial Intelligence Research and Development Strategic Plan. The Obama administration released the original plan in 2016, and the updated plan includes the original seven focus areas, such as developing effective methods for human-AI collaboration, and adds public-private partnerships as an eighth focus area. In addition, the updated plan highlights the programs the government has initiated since 2016, such as the National Institute of Health’s ongoing research in natural language processing using 96 million facts extracted from citations.
2. Predicting a Person’s Gestures
Researchers from the University of California, Berkeley have developed an AI system that can predict a person’s hand and arm gestures based on the sound of their voice. The researchers used an algorithm to extract the movements of individuals talking in 144 hours of video and used the data to train their system. The system could then predict a person’s gestures using only audio of the person speaking.
Researchers from the University of Virginia have developed a neural network that can accurately detect and distinguish between environmental enteropathy and celiac disease, which affect the gut but are difficult to differentiate because of their similarities. The researchers trained their neural network on thousands of images of biopsies, finding that the network could then distinguish between the diseases with greater than 93 percent accuracy.
Researchers from the University of California, Berkeley and the University of Southern California have developed an AI tool that can detect deepfakes, which are fake but realistic videos, about world leaders. The researchers used an open-source tool to extract the face and head movements of leaders and then used machine learning to detect the differences between a leader’s movement in deepfake and real videos. The system can detect a variety of deepfake videos with 92 percent accuracy.
5. Using Smart Speakers to Detect Cardiac Arrest
Researchers from the University of Washington have developed a machine learning tool that uses a smart speaker’s microphone to detect the warning signs of cardiac arrest and call emergency services. The researchers trained the tool on more than 80 hours of 911 calls made in King County, Washington in which the audio captured an individual suffering from agonal breathing, which is associated with cardiac arrest and is the gasping sound a person makes when they are struggling to breathe. The system can recognize agonal breathing from 20 feet away with 97 percent accuracy.
6. Reverse Engineering Recipes
Researchers from MIT have developed PizzaGAN, a neural network that can analyze a picture of a pizza and determine the type of toppings and the order someone placed them on the pizza. The researchers trained the system on 5,500 clip art-style images of pizzas and more than 9,000 images of real pizzas from Instagram. The network could help lead to AI systems that can analyze a picture of any meal and reverse engineer a good recipe for it.
7. Predicting the Development of Type 1 Diabetes
Researchers from IBM have developed an AI tool that can predict when a patient will develop diabetes by detecting Type 1 diabetes antibodies in blood. The researchers used data on more than 22,000 individuals from the United States, Finland, and Sweden to develop the tool, which found similarities between individuals with specific antibodies in their blood and the timeline of their Type 1 diabetes progression.
8. Finding Human Sex-Trafficking Victims
Investigators have used Spotlight, an AI tool from anti-trafficking nonprofit Thorn, to find more than 9,000 sex-trafficking victims. Spotlight uses Amazon’s facial recognition technology and text-processing algorithms to match faces and other clues in online sex trafficking ads with evidence. The tool has also helped investigators find more than 10,000 human traffickers.
9. Making AI Immune to Adversarial Attacks
Researchers from the Commonwealth Scientific and Industrial Research Organisation, Australia’s national science agency, have developed a method to reduce the effectiveness of adversarial attacks, which are attempts to fool AI systems by introducing malicious inputs. The researchers’ method makes small modifications to the training dataset, such as distorting the images. These distortions make the model more robust and increase its immunity to adversarial attacks.
Researchers from IBM have developed an AI system that can predict if a patient will develop breast cancer a year before the disease appears. The researchers trained their system on nearly 10,000 mammography images and the corresponding health records for patients. The system can predict breast cancer with 87 percent accuracy, which is comparable to human radiologists.