Google DeepMind Develops AI That Learns 'Physics' [Reading Science]

Understanding Object Motion Laws Like Humans Through Learning

[Asia Economy Reporter Kim Bong-su] DeepMind, a subsidiary of Google, has developed an artificial intelligence (AI) that can learn the laws of motion of objects, i.e., physics, on its own, much like a human baby. It observes the movements of objects, understands their patterns and rules, and even expresses surprise when those rules are violated.

Google DeepMind published a paper containing these findings on the 11th in the international academic journal Nature's sister journal, Nature Human Behaviour.

Developmental psychologists show videos where a ball suddenly disappears to measure the intelligence development level of children. Children generally show surprise, and developmental psychologists score this by measuring how long a baby gazes in a particular direction.

DeepMind researchers, including Lewis Piloto, created similar test software using AI. They developed a neural network called PLATO, a software that learns by collecting large amounts of data from animated videos showing simple objects like boxes or balls moving, finding patterns. PLATO stands for 'Physics Learning through Auto-encoding and Tracking Objects.' It was provided with individual images of objects appearing in videos as well as original photos. It was also designed to develop internal representations of physical properties of objects such as position and velocity.

The PLATO developed by the research team learned from videos showing simple mechanisms such as balls rolling down inclined planes or two balls bouncing together for dozens of hours. It also developed the ability to predict how these objects would behave in different situations.

After this learning process, PLATO was able to recognize and express surprise like a young child when unusual events occurred in videos, such as objects suddenly disappearing. In particular, it predicted the next movements while watching animations, and the accuracy of these predictions improved with more training. It also understood that objects do not pass through each other, maintain the same shape continuously, and move along a certain trajectory rather than disappearing magically from one place and appearing in another.

Researcher Piloto explained, "Although PLATO was not designed as a model of infant behavior, it could help develop AI that can test the learning level of infants."