Neural network can become a way to study the learning style of human infants. Inspired by the study of infant learning styles, computer scientists have developed a program that can learn simple physical rules of object motion. The research results were published in nature human behavior on July 11. Developmental psychologists track babies' eyes to test how they understand the movement of objects. For example, when children see a video of a ball suddenly disappearing, they will show surprise. Researchers quantified this by measuring the time they looked in a particular direction. Luis Piloto, a computer scientist at deepmind, a Google company based in London, UK, and his collaborators hope to develop a similar artificial intelligence (AI) test. Using animated videos of simple objects such as cubes and balls, the team trained a neural network, a software system that learns by discovering patterns in a large amount of data. This software model, called "physics learning by automatically encoding and tracking objects" (Plato), obtains the original image from the video, and is also provided with a version that highlights each object in the scene. Plato is designed to develop internal representations of physical properties of objects, such as their position and velocity. The system received about 30 hours of video training. The video showed simple mechanisms, such as a ball rolling down the slope or two balls bouncing off each other, and then developed the ability to predict how these objects would behave in different situations. In particular, it learned some patterns, such as continuity, in which objects move along uninterrupted tracks, rather than magically transferring from one place to another; Firmness, preventing two objects from penetrating each other; And the persistence of object shape. "Every step of the video predicts what will happen next. As the video deepens, the prediction will become more accurate." Piloto said. When playing a video marked with an "impossible" event, such as an object suddenly disappears, Plato can measure the difference between the video and its own prediction, thus providing a surprising measure. Pilot said that Plato was not designed as a model of infant behavior, but it may be the first step towards AI, which can test the hypothesis about how human infants learn. "We hope that this will eventually be used by cognitive scientists to seriously simulate the behavior of infants." Jeff clune, a computer scientist at the University of British Columbia in Vancouver, Canada, said that comparing AI with the learning styles of human infants is "an important research direction". Clune and other researchers are studying a way to develop their own algorithms to understand the physical world. (Xinhua News Agency)