MIT Latest Research: Teaching AI Classic Physics Understanding through Infants

2019-12-03

A research team at the Massachusetts Institute of Technology (MIT) has recently developed an artificial intelligence model that harnesses the astonishing potential of the human brain. This model demonstrates a level of understanding in physics equivalent to that of a three-month-old infant. While this may not seem like much, a three-month-old infant already has a basic understanding of how the body works. They comprehend advanced concepts such as solidity, persistence, and the fact that objects do not pass through or disappear from each other, and they can predict their movements. To study this, researchers showed infants videos that seemed to violate the laws of physics, such as objects passing through each other and reappearing on the other side. The infants displayed varying degrees of surprise in response. Kevin Smith, a researcher at MIT, said, "By the time infants are three months old, they already have some notions that objects don't blink in and out of existence. We wanted to capture and formalize this knowledge to build artificial intelligence agents that have the same cognitive abilities as infants. Now, we are modeling what is reasonable and unreasonable for infants." The main idea of the MIT team is to train AI to recognize whether a physical event should be considered surprising and then express this surprise in its output. According to MIT, a rough description of objects is input into a physics engine, which simulates the behavior of physical systems such as rigid bodies or fluids. This is commonly used in movies, video games, and computer graphics. The researchers' physics engine "propels objects forward in time," creating a series of predictions or "confidence distributions" for what will happen to these objects in the next frame. Next, the model observes the actual next frame. It captures the object representation again and aligns it with one of the predicted object representations based on its confidence distribution. If the object adheres to the laws of physics, there won't be a significant mismatch between the two representations. On the other hand, if the object performs an incredible action, such as disappearing from behind a wall, there will be a substantial mismatch. Predicting what will happen next in any given sequence of events is highly complex and requires extensive computation even for non-AI systems. Unfortunately, even AI systems are starting to yield diminishing returns under the traditional computational paradigm. So, we may have to abandon the approach of stuffing data into a black box and instead string together or use hundreds or thousands of processing units to fine-tune and organize useful outputs from artificial neural networks. Some experts believe that we need a quantum solution that can time travel or arrive at multiple outputs simultaneously and then automatically surface answers, much like the human brain. This puts us in a "Catch 22" situation because our understanding of the human brain, artificial neural networks, and quantum physics is considered incomplete. Hopefully, ongoing research in these three areas will contribute to an upward trend that boosts all ships. For now, scientists hope that artificial curiosity and the curation of "surprise" will help bridge the gap between the human brain and artificial neural networks. Ultimately, this novel, exploration-based learning approach can be combined with quantum computing technology to lay the foundation for machine thinking.

RECENT AI NEWS

RECENT AI TOOLS