This week’s list of data news highlights covers January 14-20, 2017 and includes articles about a method for diagnosing PTSD from voice recordings and an app that can detect forgeries with data from fitness trackers.
Israeli startup Joonko has developed an AI application designed to identify signs of bias in the workplace. For example, Joonko can analyze factors such as a salesperson’s success rate, experience, and tenure to determine if certain employees are being overlooked for new opportunities or regularly receiving less important tasks than others. Joonko can also notify employees when they achieve certain milestones so they can be better equipped to identify when they are unfairly passed over due to subconscious bias.
The U.S. National Highway Transportation Safety Administration (NHTSA) has closed its investigation of the fatal accident involving a Tesla using its self-driving “Autopilot” feature. NHTSA found that not only were there no defects in the technology or need for a recall, but that the technology reduced crash rates for Teslas by 40 percent. Teslas equipped with the Autopilot technology can steer themselves automatically while driving—a feature called Autosteer—based on data from a suite of sensors around the car. NHTSA found that prior to the introduction of Autosteer, the Teslas crashed at a rate of 1.3 crashes per million miles, and that with Autosteer, this rate dropped to 0.8 crashes per million miles.
Researchers at New York University’s Langone Medical Center and nonprofit research institute SRI International are developing a method for diagnosing medical conditions based on a patient’s vocal characteristics. The researchers used machine learning to analyze 40,000 vocal characteristics in voice recordings of veterans, who have a high risk of suffering from post-traumatic stress disorder (PTSD) and traumatic brain injury (TBI), and they were able to identify 30 distinct characteristics that appear to be associated with those conditions. The researchers were also able to identify a vocal characteristic, too subtle for human ears to distinguish, associated with a dramatically heightened risk of coronary artery disease.
A startup called Behavox has developed a monitoring system for financial firms that uses machine learning to analyze traders’ behavior and flag activity that suggests a trader might be involved in illicit activity. Behavox can track traders’ activities, such as the frequency of their phone calls or the times they log in to their work computers, and compare it against a database of case studies of traders found to be breaking the law. If Behavox detects suspicious behavior, it scores the behavior based on risk, maps the relationships between the employee in question and other employees, and warns managers about potential problems.
Researchers from Tel Aviv University and Ben Gurion University in Israel have developed an app that can authenticate handwritten signatures by analyzing data from motion sensors in wearable fitness trackers and smartwatches with 95 percent accuracy. The app records subtle details about wearers’ movements as they write their signature, such as arm angle and the time the signature takes to write, and compares this data against a previously-recorded, authentic signature. Though banks already use a variety of authentication technologies, this method could be useful for a wide variety of other situations, such as authenticating signatures on mail-in ballots or verifying signatures on legal contracts.
Researchers at the University of Bialystok in Poland have developed a faster memory storage technique. Normal hard drives write data by using electromagnets to change the magnetic orientation of a small patch on a spinning disc or in flash memory chips. Unfortunately, this process is energy-intensive and produces a lot of heat. By instead using rapid pulses of polarized light to change the magnetic orientation of different points in a piece of garnet, the researchers could write data 1,000 times faster than current hard drives while generating very little heat.
A number of different groups, including the nonprofit research institute OpenAI, Google’s DeepMind, and the Massachusetts Institute of Technology (MIT) are experimenting with developing machine learning systems that can develop their own learning systems. For example, Google’s Deepmind team developed software that they tasked with developing machine learning systems that could solve collections of multiple, similar problems, such as navigating different mazes, and found that their software was able to generalize and address these tasks with less training than would normally be required. This approach could make developing complex AI systems much less time-intensive for developers and more practical to apply to new scenarios.
Multiple grassroots initiatives have hosted hackathons devoted to copying and archiving federal datasets related to climate change to prevent it from potentially being lost during the Trump administration. The University of Toronto, the University of Pennsylvania, and the nonprofit Internet Archive have all hosted events to scrape data from federal websites, prioritizing data from scientific and environmental agencies.
The U.S. Department of Transportation has selected ten sites throughout the United States to serve as testing grounds for self-driving cars. Automakers will use these sites to improve and verify the safety of their autonomous vehicles before they are made widely available. The sites, located in nine different states, are designed to encourage the development of communities of practice for self-driving car safety and testing.
The U.S. Food and Drug Administration (FDA) has for the first time approved deep learning software for clinical use. The software, developed by healthcare cloud platform company Arterys, is a medical imaging tool that uses AI to help diagnose heart problems. For FDA approval, Arterys had to demonstrate that its software was at least as accurate as humans at producing test results. Not only did Arterys’ software meet this standard, but it was able to produce test results in an average time of 15 seconds, compared to the 30 minutes it typically takes human analysts.