This week’s list of data news highlights covers September 28-October 4, 2019, and includes articles about using AI to restore movement in individuals with spinal cord injuries and using AI to judge gymnastics.
1. Shrinking a World-Class AI Model
Researchers from Huawei have successfully developed a shrunken model of BERT, Google’s state-of-the-art language model that surpassed a reading-comprehension benchmark. The original Bert used 340 million data parameters and training it cost enough electricity to power a U.S. home for 50 days. Huawei’s version of BERT uses a technique called knowledge distillation in which a large AI model trains a smaller model until their outputs match. This technique helped Huawei’s model, TinyBERT, use millions of fewer parameters than BERT and be almost ten times faster. Google has also produced a significantly smaller version of BERT, but it performs slightly worse than Huawei’s model.
2. Restoring Movement After Spinal Cord Injuries
Brown University and Intel have partnered to develop an AI system that could restore movement and bladder control for individuals with severe spinal cord injuries. During the two-year project, researchers will collect motor and sensory signal data from the spinal cord, and surgeons will implant electrodes on each end of an injury to create an “intelligent bypass.” The researchers will then train neural networks to communicate motor instructions through the bypass.
3. Monitoring How Patients React to Antibiotics in Real-Time
Researchers from Imperial College in London have developed a sensor that can monitor how a patient is reacting to antibiotics in real-time. The 1.5-centimeter sensor consists of a patch with tiny needles that doctors can place on forearms and connect to monitors. In a test, the researchers found the sensors were roughly as accurate as blood tests at detecting the amount of penicillin in patients’ bodies.
Researchers from Stanford University and Intermountain Healthcare, a hospital system in Utah, have developed an AI system that can detect pneumonia in ten seconds by analyzing x-ray images. The researchers trained the system on thousands of images from the Stanford Medical Center and Intermountain hospitals’ emergency departments. The system could help patients with severe pneumonia receive treatment faster as humans can take roughly 20 minutes to diagnose pneumonia in x-rays.
5. Using AI to Judge Gymnastics
The Artistic Gymnastics World Championships in Germany will use an AI system from Fujitsu to help judges score gymnastic performances. The system uses data from sensors to track a gymnast’s movements, including angle measurements of their body, to create a 3D rendering of the performance. Judges will use the system, which can also help correct a gymnast’s form during training, when a judge or national team asks for a review of the scoring.
6. Building a Supercomputer for AI
HP has built a new supercomputer for the Lincoln Laboratory Supercomputing Center, an organization affiliated with MIT, that is designed to rapidly perform neural network operations. The supercomputer combines high-performance computing and hardware optimized for AI by using 100 Intel processors and 900 Nvidia GPU accelerators. The supercomputer will help train machine learning algorithms to support the laboratory’s initiatives, such as accelerating medical analysis and designing synthetic DNA.
7. Automating the Creation of Neural Networks
Researchers from Oak Ridge National Lab, a U.S. government-funded research lab, and Stony Brook University in New York have developed software that can quickly develop neural network architectures. The software generates millions of new models based on a researcher’s objectives, such as maximizing accuracy, minimizing training time, and minimizing the time it takes to make predictions. The researchers used the software to create models to analyze biopsy scans for cancer, and the best performing model analyzed scans for seven different types of cancer with comparable accuracy to state-of-the-art models while being 16 times faster.
8. Using Facial Recognition to Verify Identities
France is using facial recognition to verify its citizens when they create their digital ID. Individuals will use an app to take a selfie, which facial recognition will compare against their passport photo to confirm their identity. France is deleting the selfies once the enrollment process is over. The digital ID will allow French citizens to access public services like their taxes online.
9. Making Food Deliveries Indoors
Chinese food delivery company Meituan Dianping has developed autonomous delivery robots that can use both elevators and stairs to deliver food indoors. The company is testing the robots in 10 hotels and office buildings, and the robots have already made thousands of deliveries. The robots still need a human to first deliver the food to the building, however.
10. Making Deliveries With Drones
The U.S. Federal Aviation Administration (FAA) has approved UPS to run the first drone delivery service in the United States, which will allow it to expand its use of autonomous drones. UPS has been using autonomous drones to deliver medical supplies in North Carolina. The FAA still has to approve UPS operations in which drones will fly beyond an operator’s visual line of sight, which UPS has been testing in Raleigh.