|Three-month-old babies have a basic understanding of how physical things work / Photo Credit: alexnika (via Shutterstock)|
MIT researchers have unlocked the potential of the human brain to create an AI model that is capable of understanding physics as good as some humans, particularly three-month-old babies, said Tristan Greene of The Next Web, a website dedicated to new technologies and start-ups in Europe. Three-month-olds have a basic grasp of how physical things work, as they can understand advanced concepts like solidity and permanence and predict motion.
With that, MIT researcher Kevin Smith and his colleagues showed infants “videos of objects acting the way they should such as passing behind an object and emerging on the other side,” while others completely broke the laws of physics. The scientists learned that the infants show “varying levels of surprise” when objects don’t act the way they should. Smith explained, “We wanted to capture and formalize that knowledge to build infant cognition into artificial-intelligence agents.” He added that they are now getting closer to how models can differentiate basic implausible or plausible scenes in a human-like fashion.
The idea here is to train AI to recognize whether a physical event should be perceived as surprising or not, and express that surprise in the model’s output. A physics engine is fed with descriptions of coarse objects, according to Rob Matheson of MIT News, a media platform dedicated to highlighting the achievements of the MIT community. The said engine is used for video games, computer graphics, and films, stimulating the “behavior of physical systems, such as rigid or fluidic bodies.”
“The researchers’ physics engine ‘pushes the objects forward in time,’ [per paper coauthor Tomer Ullman],” creating a range of predictions or “belief distribution” for what will happen to the objects in the next frame. The model captures the object representation as soon as it reads the next frame, aligning the object to “one of the predicted object representations” from the model’s belief distribution. If the object did not break the laws of physics, it’s a match. Otherwise, it will be a major mismatch.
Hopefully, this study could be utilized with quantum computing technology to serve as the foundation for “thinking” machines.