New AI From MIT Can Intuitively Understand Some Physics Laws

Artificial intelligence is often considered a cutting-edge tool that effectively automates human tasks - but it also helps people to better understand themselves. Recently, researchers at the Massachusetts Institute of Technology (MIT) have designed a new AI product that can serve this goal. 

Named ADEPT, this AI can intuitively understand some basic physics laws. Through observing an object in a video, the AI model can predict the object’s next actions based on physical laws. It can record actions that may disappear then or move immediately. The creators behind ADEPT suggested that this model could assist future researchers in creating smarter AI, and help us approach the way infants perceive the world through their eyes. 

ADEPT will go through two steps in its information processing. The first step is to examine an object by determining its shape, velocity, and posture in a general way through the nearest geometric model of the object, rather than going into analysis with details for each aspect. An advantage of this step is that the AI ​​model can predict the movement of many different objects, not limited to what it has been trained. This is similar to the characteristics of babies, as they only care about how objects move, not too much about specific physical properties. 

The second step will involve physical systems. The AI ​​model will gather the data collected and simulate the actions of the object based on physical laws. There will be some predictions for the next actions of the object, and the model will help compare with the frames in the videos. When a difference is detected, it sends a signal. The stronger the signal, the stronger the model response to what had just happened. This approach is quite similar to the way game developers are doing to reproduce physical appearance in real-world games.

In the future, the team will continue to study how children view the world around them and integrate those findings into improving the AI ​​model.