DeepMind Explores the Field of Intuitive Physics to Achieve Common Sense in AI

DeepMind

DeepMindDeepMind researchers develop a DL model to learn intuitive physics inspired by developmental psychology

Although recent years have seen striking advances in artificial intelligence, it is increasingly remarked that most of this progress is limited to narrow domains. It has made astounding strides. Modern deep learning systems are particularly challenged by the “common sense” information that directs prediction, inference, and behavior in commonplace human contexts. In the present work, DeepMind researchers focus on one particular domain of common-sense knowledge: intuitive physics, the network of concepts.

This ability, dubbed “intuitive physics,” seems extremely trivial on the surface. It underlies reasoning about the properties and interactions of macroscopic objects.  To tackle the trickier challenge presented by intuitive physics, the researchers developed a model called PLATO, which stands for Physics Learning through Auto-encoding and Tracking Objects to focus on whole objects instead of individual pixels. Intuitive physics is fundamental to embodied intelligence, most obviously because it is essential to all practical action

 

DeepMind Gave an AI Intuitive Physics:

The DeepMind research team was frequently inspired by developmental psychology, where the acquisition of intuitive physics knowledge has been a major area of study to explore richer common sense physical intuition in AI systems. They also created a means of testing the model that is akin to the methods used to assess cognition in human infants. They trained PLATO on about 300,000 videos so that it could learn how an object behaves: a ball falling, bouncing against another object, or rolling behind a barrier only to reappear on the other side.

Babies rapidly develop this ability by soaking up data from their external environments, forming a sort of “common sense” about the dynamics of the physical world. This intuitive physics model by integrating the same inherent knowledge that developmental psychologists think a baby is born with into an AI system. Additionally, researchers used developmental psychology to address the issue of behaviorally determining whether an AI system is knowledgeable about intuitive physics.

In developing behavioral probes for research on children, developmental psychologists have based their approach on two principles. First, the foundation of intuitive physics is a collection of distinct ideas that can be distinguished, operationalized, and individually investigated. The second rule that developmental psychologists employ when examining physical concepts is that having a physical concept is equivalent to developing a set of expectations about how the future might play out.

The goal was to have PLATO understand what violates the laws of intuitive physics based on five fundamental concepts: object permanence, solidity, continuity, unchangeableness, and directional inertia. PLATO isn’t a digital replica of a three-month-old baby and was never designed to be. Critically, the Physical Concepts dataset also includes a separate corpus of videos intended as training data. The DeepMind research team believes it would be exciting to expand on the current modeling work to engage with important developmental psychology issues even more directly.