Reverse-Engineering Physical Common Sense
-----
People have no trouble judging whether a stack of dishes is stable, pouring coffee without spilling, or judging if a box is full based on how it sinks into a sofa. We find these capabilities so simple we call them common sense, yet they far outstrip the abilities of state-of-the-art AI. In this talk, Dr. Smith will discuss his research towards reverse engineering this physical common sense: studying the cognitive processes underlying these intuitions about the world, and using these insights to develop human-like AI. First, he will present a cognitively-inspired AI system that parses a scene into objects, tracks them through a video, and predicts future events, and show that this system better explains human predictions than internet-scale deep learning models. Next, Smith will share research on the rational approximations that allow people to reason about the world under cognitive capacity limitations: that we, for instance, do not represent entire scenes but instead build up our representations incrementally to support our reasoning. Finally, Smith will discuss future directions to build these rational approximations into AI, pointing towards models of generalized physical common sense that can serve both as candidate models of human cognition and useful AI systems.
-----