This week’s list of data news highlights covers November 3 – 9, 2018, and includes articles about an AI system that can dress itself and new AI-focused healthcare centers opening in the United Kingdom.
1. Mapping Mines and Tunnels With Drones
Australian startup Emesant has developed software to enable drones to autonomously map mines and tunnels with drones, which could help reduce the need for humans to enter dangerous environments. The software uses LIDAR, sensors, and GPS to generate 3D maps of the mines and tunnels. Emesant has already used their system to map a mine 2,000 feet below the surface of Western Australia.
2. Parking Cars with AI Based on a Worm’s Brain
Researchers at the Australian university TU Wein have developed an artificial neural network modeled after the C. elegans worm capable of parking a small remote control car. The C. elegans worm is the only organism to have its entire nervous system and brain mapped out, serving as a guide for the researchers, who built their neural network with just 12 artificial neurons modeled after the worm’s neural circuit that controls its physical reflex to touch. Using this model, the researchers were able to train the system to parallel park a small car equipped with proximity sensors.
3. Teaching AI to Put on a Shirt
Georgia Institute of Technology researchers have developed an AI system that controls a humanoid model that can put on a shirt. The act of putting on a shirt, though straightforward, requires multiple different kinds of movements with different forces in a specific sequence, which can be challenging for AI systems to learn. The researchers used a technique called reinforcement learning to reward or penalize the system if, for example, it ripped the shirt when putting it on, and trained it on each discrete step involved. This approach could make it easier to develop robotics systems capable of performing a variety of different, complex movements.
4. Fighting Fake Science with Data Science
Researchers are increasingly using analytics techniques like AI to scrutinize suspicious scientific research and identify if it uses falsified data. For example, a company called Resis specializing in verifying research integrity has developed software that can detect graphs that are too similar to graphs that appeared in other research, suggesting the authors copied the graph rather than created it themselves. Additionally, a machine learning researcher at Syracuse University developed an algorithm to identify duplicate images in hundreds of thousands of biomedical research papers, finding that 9 percent of images were suspiciously similar.
5. Dreaming Up New Materials with AI
Massachusetts startup Kebotix has developed a system that uses AI and robotics to formulate new chemical compounds that could be valuable in materials research. The system uses machine learning to analyze molecular models with desirable properties and generate predictions of new compounds that could exhibit the same properties. Then, the system uses a robotic arm to test its predictions by combining different chemical samples into a dish while analyzing the results, which are then fed back into the system to allow it to continuously improve its predictions.
6. Using AI to Trace The Source of Food Poisoning
Google and Harvard University researchers have developed a machine-learning model called FINDER (Foodborne Illness Detector in Real Time) to identify restaurants likely to cause food poisoning. FINDER uses anonymous Google search query and location data to compute the number of individuals who visited a restaurant and then searched for terms indicative of food poisoning. Restaurant inspectors in Las Vegas and Chicago used FINDER and found that establishments flagged by FINDER as high-risk were over three times as likely to be deemed unsafe upon inspection.
7. Predicting the Right Medical Treatment
Massachusetts Institute of Technology and Massachusetts General Hospital (MGH) researchers have developed an AI model that can help doctors treat sepsis, a blood infection that kills nearly 270,000 U.S. patients every year. The researchers used data from 186,000 patients who visited MGH’s emergency room from 2014 to 2016 to train and test the model’s ability to predict when doctors should administer vasopressor medications such as dopamine, which can be effective but cause health problems if administered at the wrong time. The AI model uses data features such as a patient’s respiratory rate, blood pressure, and amount of blood their heat pumps in each beat to predict the correct treatment timing over 80 percent of the time.
Researchers from Hong Kong Baptist University and Chinese technology company Tencent have developed a new technique that trains AI machines faster without losing accuracy. The researchers used simpler computational methods and combined smaller pieces of data into larger ones to train two popular neural networks, AlexNet and ResNet-50, faster than ever before—by four and nearly seven minutes, respectively..
Chinese state-run news agency Xinhua has developed two AI anchors to deliver broadcast news. The agency worked with Chinese search engine Sogou to create the realistic-looking virtual anchors that use AI to mimic realistic speech patterns and facial movements. The AI anchors must be fed scripts, but could allow the agency to reduce production costs and have an “anchor” delivering the news around the clock.
10. Bringing AI-Driven Healthcare to the United Kingdom
The United Kingdom’s National Health Service (NHS) has announced it will open five health clinics in 2019 at universities and NHS facilities that specialize in AI-powered diagnostics. In particular, the centers will focus on advancing medical imaging and analysis to improve outcomes and enable NHS staff to spend more time on direct patient care. The centers are funded through the United Kingdom’s Industrial Strategy Challenge Fund and run in partnership with companies including GE Healthcare, Siemens, and Leica.
Image: Dmadeo.