From driverless vehicles to weapons systems, artificial intelligence (AI) models are being trusted with an awful lot of responsibility these days, so you’d like to think the technology has some idea of what’s going on. Fortunately, we can all now rest assured thanks to the whizzes at DeepMind, who have created the first ever AI with a grasp of physics comparable to that of a human baby.
Writing in the journal Nature Human Behaviour, the researchers explain how we develop “intuitive physics” within the first months of life, quickly coming to understand certain fundamental laws governing the material world. For example, infants tend to comprehend the concepts of “permanence” – whereby objects don’t simply vanish – plus “solidity” and “continuity”, referring to the inability of objects to pass through one another or to suddenly alter their trajectory through time and space.
However, the authors go on to state that “current artificial intelligence systems pale in their understanding of intuitive physics, in comparison to even very young children.” To help the bots catch up, the team turned to the field of developmental psychology to develop an AI that is capable of learning in the same way as a baby.
For instance, by the age of about three months, human infants are capable of showing surprise when an object disobeys one of the three pillars of our intuitive physics. This ability is known as the violation-of-expectation (VoE) paradigm, and provides the inspiration behind the new AI.
Called PLATO – standing for Physics Learning through Auto-encoding and Tracking Objects – the deep-learning system was trained on a series of videos of balls moving through space and interacting with one another. The video dataset was specifically designed to represent the concepts of permanence, solidity and continuity, as well as two extra concepts known as “unchangeableness” and “inertia”. These relate to the fact that objects do not suddenly alter their basic characteristics or disobey the laws that govern speed and direction.
When PLATO was later shown videos of scenarios that contravened any of these five tenets, it successfully reacted with a VoE signal. “After training PLATO on videos of simple physical interactions, we found that PLATO passed the tests in our Physical Concepts dataset,” explained study author Luis Piloto in a statement.
“By varying the amount of training data used by PLATO, we found that PLATO could learn our physical concepts with as little as 28 hours of visual experience.”
Impressively, the AI could even identify transgressions of the laws of physics when looking at a separate video dataset featuring objects that it had never seen before. “PLATO passed, without any re-training, despite being tested on entirely new stimuli,” says Piloto.
This breakthrough certainly bodes well for the future of AI, since, As Piloto points out, “if we’re to deploy safe and helpful systems in the real world, we want these models to share our intuitive sense of physics.”
Obviously there’s still a way to go, but baby smarts ain’t a bad start.