This week’s list of data news highlights covers November 24–30, 2018, and includes articles about astronauts getting an AI assistant and a startup using autonomous drones to replant forests after fires.
The Nevada Highway Patrol (NHP) used artificial intelligence to reduce the number of crashes by 17 percent on a portion of Interstate 15, one of the state’s busiest highways. NHP partnered with Tel Aviv startup Waycare to implement a system that analyzes data from cameras and sensors to predict which areas are at high risk for accidents. The predictions let NHP take preventive actions, such as displaying its vehicles in highly visible areas to slow down speeding drivers. And when accidents do occur, the data helps NHP identify them up to 12 minutes faster.
Uber researchers have developed algorithms that can outperform most humans at playing the 1980s video games Pitfall! and Montezuma’s Revenge. AI had previously struggled at performing well at both games because they do not award players for positive behaviors until far into the games, making it difficult for AI to associate a behavior with a positive reward. The researchers solved this dilemma by developing algorithms that can remember their past performances and return to a task or area to attempt to generate better results. Such an approach could be useful in robotics, where robots often will not get immediate positive feedback for positive behavior.
Volvo will begin using six autonomous trucks to transport limestone through three-mile tunnels in a Norwegian mine in 2019. Volvo is currently testing the vehicles, which use preset routes that should make them easier to deploy. Rather than selling the trucks, Volvo will be providing the service to local mining firm Brønnøy Kalk AS, which will pay Volvo per ton it delivers.
U.S. startup WaveOne has developed a video compression algorithm that uses deep learning to significantly outperform existing standards, which could help reduce the time it takes to download videos. Video compression works by replacing redundant descriptions, such as the color or location of an object in each frame, with shorter descriptions that can still accurately reproduce the video. WaveOne’s algorithm needs less information than traditional compression algorithms to accurately predict the location of a moving object in each frame, allowing it to make video files smaller than traditional compression algorithms.
Academic journals are increasingly using AI tools to make peer reviewing easier. For example, dozens of publishers are piloting StatReviewer, a peer-review platform that identifies fraudulent statistics in papers by analyzing information such as sample sizes and baseline data. In addition, peer-review platform ScholarOne is working with UNSILO, which provides AI tools to publishers, to alert editors to the key statements in a paper. UNISILO’s AI also highlights if a paper’s claims are similar to those of past papers to identify plagiarism and put a paper in context.
Seattle-based startup DroneSeed is replanting burned forests with drones and AI. DroneSeed uses autonomous drones equipped with LIDAR and cameras to fumigate areas burned by forest fires, so invasive weeds do not grow in the place of trees, and to collect imagery of the landscape. AI then analyzes the landscape to identify the best places to plant new seeds, which the firm customizes for each location. With the United States losing over 7 million acres a year to forest fires, the combination of drones and AI could reduce the number of humans who do the difficult and time-consuming work of spraying burned forests and planting new trees.
An AI system called the Crew Interactive Mobile Companion (CIMON) had its first conversation with a crewmember while aboard the International Space Station (ISS). CIMON interacted with German astronaut Alexander Gerst, successfully recognizing the astronaut and providing instructions for an experiment with crystals. Developed by the German Aerospace Center, Airbus, and IBM, the medicine-ball-sized CIMON freely floats about the ISS and communicates with IBM’s Watson AI on Earth to answer questions.
Boston-based startup Lightelligence is developing computer chips that use light instead of electrons to perform the mathematical computations involved in deep learning. Photons are significantly faster than the electrons used in today’s computer chips, which could allow AI algorithms to run faster while also not overheating a chip’s circuitry. In practice, the speed of the chip will depend on how fast it can interact with traditional components of a computer, such as its memory.
The U.S. Office of the Comptroller of the Currency, the Federal Deposit Insurance Corporation, and the Federal Reserve have proposed relaxing real-estate appraisal requirements to allow a majority of U.S. homes to be sold without a licensed human appraisal, which could increase the number of appraisals based on algorithms. Under the proposal, individuals could buy or sell any home under $400,000 without an appraisal from a licensed human, which can cost from nearly $400 to $900. Valuations done by services who use algorithms can cost less than $100.
The Massachusetts Institute of Technology and Facebook have developed a machine learning method for providing addresses to many of the four billion people on earth who do not have one. The researchers used algorithms to identify roads from satellite imagery and connect the roads into a network of communities. The system then labeled the densest areas as city centers and numbered and lettered streets according to how far they were from the city center. This approach created addresses for more than 80 percent of a random sampling of unmapped but populated areas. The researchers now want to work with nonprofits to use the system to provide people addresses.