This week’s list of data news highlights covers February 10-16, 2018, and includes articles about an AI system that can detect small earthquakes and a genetic analysis technique that could indicate time of death.
London startup ASI Data Science, working for the UK Government, has developed a machine learning system that can detect 94 percent of ISIS propaganda videos uploaded online with a false positive rate of just 0.005 percent. ASI trained the system on thousands of hours of ISIS videos to teach it to identify characteristics, such as logos and metadata, indicating a video was made by ISIS for propaganda purposes. When the system detects a video likely to belong to ISIS, it flags it for human review and potential removal.
Researchers at Harvard University have developed an AI system that can increase the sensitivity of seismographs to detect 17 times more earthquakes than traditional methods. The system can analyze seismograph data and differentiate between normal geological activity, known as ambient seismic noise, and earthquakes that might be masked by this activity due to them being very small or far away. The system makes seismographs sensitive enough to detect earthquakes of magnitude zero or minus one, which typically produce readings undetectable to the human eye.
Researchers at the University of South Carolina and Zhejiang University in China have developed an algorithm that can determine if a person using a smartphone’s touchscreen is an adult or a child to help block children from accessing inappropriate content or making online purchases. Because children have smaller hands than adults, they typically touch less of the screen and make shorter swipes, and also tend to swipe more slowly than adults and take longer to switch from swiping to tapping. The algorithm can identify if a user is a child or adult by having them play a simple game involving swipes and taps. The algorithm is 84 percent accurate after just one swipe, and 97 percent accurate after eight swipes.
Image hosting company Gfycat has developed an AI system that can fight back against the recent trend of using AI software to face-swap celebrity faces into porn videos, known as “deepfakes.” When a user uploads a GIF to Gfycat, the company uses AI to automatically search the Internet for higher-resolution versions of the same GIF and tag GIFs based on faces it recognizes to make it easier for users to find content. Gfycat has modified this system so that when it can only make partial matches to other GIFs, it will mask faces in the GIF and attempt to match bodies and backgrounds to other GIFs, which could indicate that the clip was altered.
Nebraska’s state prescription drug database is now tracking every prescription drug sent or ordered in the state in an effort to combat addiction and give health care providers better data about what patients are prescribed. States typically only track prescriptions for drugs that are particularly dangerous or have a high risk of abuse, such as opioids. Nebraska is the first state to include every prescription in its drug monitoring database, enabling doctors to be better informed about their patients and expose doctors and pharmacists that overprescribe dangerous drugs.
Alibaba has developed AI software that can recognize pigs on a farm based on a mark placed on their bodies and monitor changes in their activity. The software links each pig with data about its age, weight, and other characteristics, tracks its movement to assess its fitness, and can monitor sounds that could indicate when a pig is sick, such as coughing. The software can also use this data to predict which pigs are the most likely to produce healthy offspring.
Researchers at Waseda University in Japan have developed a deep learning system that can fill in missing portions of photographs with realistic-looking imagery. The researchers created a dataset of 8 million pictures of landscapes, human faces, and other subjects and generated multiple versions of each image with random portions blanked out to train the system to predict what is missing and generate replacement imagery that is consistent with its surroundings. The researchers were also able to use the system to produce 3D imagery from 2D photographs.
Google has worked with two companies that oversee 911 call centers in the United States to trial a method for sharing smartphone location data with emergency responders more accurately than wireless carriers. In a test involving fifty 911 call centers in three states, Google’s location data was more accurate in 80 percent of calls, with an average estimate radius of 121 feet, while carrier data had an average radius of 522 feet, and reached call centers faster than carrier data.
The Allen Institute for Artificial intelligence has developed an open-source simulator called AI2-THOR to serve as a virtual training ground for AI systems to learn about how to interact with household objects. AI12-THOR includes 120 near-photorealistic scenes of bedrooms, kitchens, bathrooms, and living rooms filled with interactive objects such as microwaves and couches, and uses realistic physics. By creating a realistic training environment, an AI system could more easily adapt things it learns in a virtual environment, such as that it could knock over a chair or slice an apple, to the physical world.
Researchers at the Centre for Genomic Regulation, a genomics research organization in Barcelona, have developed a machine learning system that can analyze gene activity in human tissue samples and predict when the person died. Genetic activities, such as the regulation of metabolism and stress responses, can increase or decrease in different ways after a person dies and can continue for several days. The researchers trained their system on tissue samples from 399 people labeled with their time of death and were able to identify predictable patterns in how different genetic activities changed at different intervals after death, which could eventually be used to help identify when a person died.