The study of physics is something we, as humans, understand early on in our life. Now, researchers at MIT have tapped into the methods of processing a bay’s brain uses to work out physics to teach AI about it as well. Babies as young as three months old understand concepts such as object permanence and the laws of physics that ensure physical bodies don’t teleport or overlap one another. Teaching AI these physical rules of the universe sets a framework in place for its future processing.
Near Human Cognition of Physical Rules
Artificial intelligence needs rules to function correctly. Once it’s given a set of rules, it can process incoming data based on those rules. Just like any other simulation, AI can apply those rules to data and spit out results. However, it can also learn what it should be looking for based on those results and change its input the second time around to get the result it wants. Understanding physics like a human allows the AI to visualize a scenario the way a human would, but with far more accurate output.
Teaching the AI About Surprise
The researchers at MIT who engaged in developing the AI’s cognition were focused on teaching it how to tell what a ‘surprising’ result was. Using what they termed a ‘belief distribution’ – a series of inferences based on physical rules which the AI knows – the system would track an object and predict what would happen to it in the next frame. If the item did not obey the rules that the AI was aware of, then it would express surprise. Teaching an AI to understand classical physics by hard-coding the information would be a nearly impossible task. This way, the AI teaches itself about the physical world based on its ‘surprise.’
Setting the Stage for More Complex Understanding
At the moment, despite the dependence on mathematics to understand the world around us, computers and AI are notoriously absent from fields like quantum mechanics research. It may be decades before we see AI get to the point of being able to understand quantum physics, but for now, researchers are happy with teaching the system to understand surprise and learn from it.