Inspired by the way babies learn, Deep Mind’s computer scientists have developed a program that learns the simple physical rules of how objects behave. The research report was published in the journal Nature Human Behaviour. Babies express surprise when they see images that violate the rules of physics, such as a video where the ball suddenly disappears. But AI lacks the ability to understand such behaviors. Luis Piloto and colleagues developed a software model called Physics Learning through Auto-encoding and Tracking Objects (PLATO) that learns the simple rules of physics like a baby. The research team trained PLATO by showing it a number of videos depicting simple scenarios, such as a ball falling to the ground, rolling behind other objects and reappearing, with many balls bouncing back and forth. After training, the researchers tested PLATO by watching videos that sometimes contained impossible scenes. Like young children, PLATO showed “surprise” when he saw “impossible scenarios,” such as objects passing through each other without interacting. PLATO achieved the above learning effect after only watching 28 hours of video. These results have major implications for both AI and human cognition research. The research team said that the model can learn a variety of physical concepts, and reflects the characteristics consistent with the findings of developmental psychology, and PLATO can be used as a powerful tool for studying how humans learn intuitive physics, while also showing that object representation is important for humans. Understanding the world around you plays an important role.
This article is reprinted from: https://www.solidot.org/story?sid=72103
This site is for inclusion only, and the copyright belongs to the original author.