This week’s list of data news highlights covers May 12-18, 2018, and includes articles about networked sensors that monitor cows and an AI system that could free people from boring conference calls.
Researchers at DeepMind have developed an artificial neural network that uses a reward mechanism modeled after how dopamine works in the brain to improve learning. The researchers designed a reward prediction error, which is responsible for optimizing an algorithm in response to positive and negative feedback, to imitate the effects of dopamine for a recurrent neural network, which learns through trial and error. In a series of tests, the researchers found that this approach was particularly well-suited for meta-learning—the process of quickly learning trends or rules from examples and applying them to solve problems.
A startup called Livestock Labs has developed a system for monitoring livestock activity and health by using sensors embedded underneath a cow’s skin. The system uses a Bluetooth-enabled device called EmbediVet that logs data about a cow’s temperature, chewing frequency, activity levels, and other factors that could be relevant to a cow’s health, and transmits this data to base stations throughout a farm. Livestock Labs is using this data to train an AI system that can provide a greater level of insight into a cow’s health, such as when it is about to give birth or is beginning to fall sick.
Researchers at the University of California, Santa Cruz, have developed an AI system called MarioGAN (generative adversarial network) that can generate game levels for Super Mario Bros. with little human input. While computer generated levels are common in video games, they typically require humans to program specific parameters to ensure the level is playable. MarioGAN creates levels by analyzing existing levels of the game and using two neural networks that generate and evaluate levels to determine if they are playable. The researchers were also able to teach MarioGAN to design levels with different levels of difficulty and make levels that get progressively more challenging.
Louis Martinez, a professor at the University of Chicago, has published an analysis of 25 years of satellite data that suggests China, Russia, and other authoritarian countries routinely falsify their gross domestic product (GDP). Martinez analyzed changes in nighttime light levels in satellite imagery, which can serve as a measure of economic activity, and found that in free democracies such as the United States and Canada, a 10 percent increase in average nighttime light intensity in a year correlated with a 2.4 percent increase in GDP for that year. However countries ranked as the least free, according to watchdog organization Freedom House, reported between a 2.9 to 3.4 percent increase in GDP for the same light intensity increase, suggesting those countries fake their GDP reports to maintain political power.
A startup called Analytic Flavor Systems has developed an AI platform called Gastrograph that can predict flavor preferences for narrow segments of the population. Many food and beverage companies rely on sensory science, which focuses on measuring responses to smell and taste, to help develop new products, however this often involves crude data gathering methods and can only identify broad trends. The Gastrograph smartphone app allows users to identify 24 categories of flavor qualities they detect in a food, such as “floral” or “gamey,” and indicate their preference for that food, while Gastrograph logs their age, ethnicity, gender, and other personal data as well as sensor data from their smartphone, including temperature and location. Analytic Flavor Systems claims this data allows Gastrograph to predict flavor preferences for very narrow demographics, which could allow food and beverage companies to develop more specialized offerings.
Teleconferencing company Dialpad has developed a service called VoiceAI that can listen in on conference calls and transcribe the conversation in real-time. VoiceAI can also provide sentiment analysis of the conversation, as well as automatically identify important moments in the call and create action items based on tasks participants discussed.
Researchers at the University of Illinois Urbana-Champaign and Intel have developed an AI system that can reconstruct dark images to make them brighter and clearer. The researchers trained their system on two datasets of 5,000 photographs taken in low-light, one of which was images that were deliberately too dark, and the other was of the same images with a longer exposure time, allowing the camera’s sensor to collect more light and take a better picture. The system is capable of reproducing a low-light image with the equivalent of up to 300 times the original exposure length but without the flaws that photo editing software creates when lightning images.
Neuroscientists at Princeton University have developed over 3,000 detailed maps of neurons with the help of video game called Eyewire that allows members of the public to parse through data about neurons and color-code portions of cells. Eyewire launched in 2012 and over 265,000 people have color-coded over 10 million sections of cells spanning 4.5 microns each. Initially Eyewire players could map a neuron over a period of weeks, but now the game is popular enough that players map several neurons per day. Eyewire’s maps are a valuable resource for neuroscientists as they provide unprecedented visualizations of individual neurons and types of neurons combined with data about their function.
Researchers at the Massachusetts Institute of Technology have developed a virtual training system for autonomous drones called Flight Goggles that allows drones to fly in a safe, open environment while they perceive a different, simulated environments. Flight Goggles relies on motion-capture technology that can watch a drone as it flies and feed this data to the drone’s sensors so it believes it is moving around a virtual environment. This approach allows researchers to train autonomous drones without having to build a large variety of complex environments or risk causing damage to drones or their surroundings.
Snohomish County police in Washington arrested a 55-year-old man for a double murder he committed in 1987 thanks to GEDmatch, the same public genealogy databased California police used to identify and arrest the notorious Golden State Killer. A company called Parabon NanoLabs, which conducts genetic analysis for law enforcement agencies, uploaded sequenced DNA from the murder scene to GEDmatch and found several relatives of the suspect, which allowed authorities to track down the suspect.
Image: Leszek Leszczynski.