Neural networks helps mimic babies' learning processes.
Do you ever wonder how we progressed as infants from learning the physical rules in the world around us? For example, how did we know that a ball rolls across the table, but a cube stays still. Do you ever see babies acting surprised when you show them a magic trick, for example, an object disappearing in front of their very eyes? Well, this is because our minds can detect patterns of processes, such as motion, and babies learn that as they grow up to be able to predict simple physics and motion.
Computer scientists created a program to imitate these processes –– a program that can pick up physical rules about the behavior of objects –– the violation of these rules brings out emotions of surprise. Developmental psychologists work to track babies’ gaze to understand the motion of objects. Google-owned company DeepMind developed a similar software for AI in which they trained a neural network. Neural network is a software that is able to learn by detecting patterns in large data using animated videos of simple objects. The name of this software is Physics Learning through Auto-encoding and Tracking Objects (PLATO), and it works by feeding raw images from the videos and detecting the objects’ positions and velocities.
This system was trained by various videos showing simple mechanisms of objects to show the difference of behavior of objects in different environments. For example, a video of a ball rolling down a hill versus a video of a ball on a flat surface. PLATO was able to learn patterns of motion and gravity, along with patterns of motion such as continuity. The more videos that were analyzed, the more the software was able to accurately predict the next moves.
The most interesting art of this software was its resemblance to the learning processes of babies. For example, when PLATO was shown an object in motion paired with unlikely/impossible events, such as the ball disappearing in this air, it was able to portray the measure of surprise, just like infants would do.
Scientists are hoping to use this software to detect and analyze babies' behaviors in the future but this is just the first step. Comparing AI to humans in research helps us analyze our learning processes much better
