This week’s list of data news highlights covers July 13-July 19, 2019, and includes articles about AI systems predicting conflicts and detecting child abuse.
IBM and Rensselaer Polytechnic Institute in New York have developed an immersive classroom that uses AI and virtual reality to teach Mandarin Chinese. The classroom, a 360-degree environment that places students in scenes in Beijing or a Chinese restaurant, uses sensors to track where a student moves or points and speech recognition algorithms to compare students’ pronunciation to native speakers. Studies have found that interactive learning environments can increase language understanding and retention.
Researchers from Google have developed a neural network that can interpret speech of people with speech disorders. The researchers trained the network on 30,000 hours of individuals speaking different dialects and using audio from individuals with speech impediments. For example, the researchers trained the network on 15 hours of audio of a deaf individual speaking, which reduced the word error rate for Google’s speech recognition engine from 89 percent to 25 percent.
Researchers from the Water, Peace, and Security Partnership, a collaboration between several organizations to design tools that can identify locations where water insecurity can lead to conflicts, have developed a tool that can predict conflicts up to a year in advance. The system uses data from NASA and the European Space Agency satellites to monitor water resources globally and combines this data with social, economic, and demographic information to predict at-risk areas. The tool predicted more than 75 percent of the water-related conflict in Mali’s Inner Niger Delta.
Researchers from the University of Potsdam in Germany have developed an AI system that can identify people by their eye movements. The researchers trained their system on two datasets, one which included eye-tracking data from people reading and another where individuals watched a dot randomly move on a computer screen. The researcher’s system achieves greater than 99 percent accuracy after analyzing ten seconds of video.
Researchers from MIT and the Polytechnic University of Milan have developed an AI tool that uses a generative adversarial network (GAN) to create Renaissance-style paintings of individuals. The researchers trained the GAN algorithm on 45,000 Renaissance portraits and the tool only needs one image of a person to create a portrait. Interestingly, the tool likely will not create a portrait of a person smiling, even if the input image is of a person smiling, reflecting the fact that most Renaissance artists resisted painting overt expressions because they thought they distorted subjects’ faces.
UK police are using an AI system to detect child abuse online. Researchers labeled images from previous police investigations by the type of offense to train the AI system, which can match faces and locations across a database of 13 million images. The system has helped the police increase the number of images they process from 18 images per minute to 200 images per minute.
Google is working with the Trevor Project, a California-based nonprofit that offers crisis counseling to LGBTQ teenagers, to develop an AI system that can predict a teenager’s suicide risk. Google will train the system using data from the beginning of a teenager’s conversation with counselors and the risk assessment counselors complete after each conversation. The system could help the Trevor Project identify high-risk individuals and intervene if necessary.
Intel has developed AI chips that can process certain types of data up to one thousand1,000 times faster than more general-purpose chips such as graphical processing units. The chips attempt to mimic the functioning of neurons and synapses in the brain and are already improving the performance of prosthetic limbs and the accuracy of digital maps for autonomous vehicles.
Researchers led by a doctor from Pennsylvania State University have shown that a machine-learning algorithm can predict patient rehospitalization and mortality following a percutaneous coronary intervention (PCI), a procedure that opens blood vessels in the heart. The researchers had the algorithm analyze demographic and clinical data concerning nearly 12,000 patients who underwent PCIs between 2004 and 2013. The researchers found that the algorithm had a 90 percent probability of correctly predicting 30-day readmissions and an 88 percent probability of correctly predicting death within 180 days.
Researchers from DeepMind have developed an AI system that can create coherent and novel 256 x 256-pixel videos. The researchers trained the system’s generative adversarial network using data from Kinetics-600, a dataset of 500,000 10-second long YouTube clips. The system was able to begin creating videos with recognizable moving objects, such as a person skiing on a mountain, after as little as 12 hours of training.