A new study draws on developmental psychology theory to teach an artificial intelligence some fundamentals of physics.
Shared in the journal Nature, the study was carried out by a team of researchers from the company DeepMind. They created a program capable of learning simple physical rules about the behavior of objects.
Learn “common sense”
While AI has made amazing progress in recent years, the most advanced systems still struggle with “common sense.” Common sense that guides prediction, inference and action in everyday scenarios. DeepMind scientists admit, however, that this intuitive physics is “fundamental to embodied intelligence,” because “it is essential to all practical interaction,” they write.
To carry out their test, the research team led by Luis Piloto was inspired by the psychology of child development. Developmental psychologists study how babies follow the movement of objects by following their gaze. When a video is shown of a ball suddenly disappearing, for example, young children express surprise. The researchers thus measure their surprise by the duration of their gaze in a given direction.
This is precisely what the DeepMind team of scientists set out to find out, by training an AI dubbed “PLATE” on videos of simple objects, such as rolling and colliding balls.
Five key physical concepts
To do so, they focused their attention on five physical concepts considered “core” in the developmental psychology literature: continuity, object persistence, solidity, immutability, and directional inertia. “Each video with a violation of these physical principles is matched with a corresponding video that provides a baseline that fits the physical principles while maintaining precisely the same single-frame statistics for all videos,” the researchers explain.
The dataset also includes a separate corpus of videos intended to serve as training data.
Using this process, the researchers wanted to build a model capable of learning what they call “intuitive physics” like a child discovering the world. The method used to measure knowledge of a physical concept is called the “expectations violation paradigm,” the researchers say.
28 hours of videos
The researchers observed learning effects after just 28 hours of videos. Definite progress, even if they acknowledge that “the range of object and event types in our data set remains narrow compared to those found in the real world.”
By domino effect, the researchers also wondered about the potential implications of their work for developmental psychology itself. “This topic should be approached with some caution, as the model we present is not intended to provide a direct model of physical concept acquisition in children. However, we believe that there are several ideas that can be useful for developmental science. First, our modeling work provides a proof of concept demonstrating that at least some core concepts of intuitive physics can be learned through visual learning.