Seminar, Hosted by Cognitive Computational Neuroscience (CCN)
Intuitive Physical Reasoning in the Human Brain
Massachusetts Institute of Technology
Abstract
Visual scene understanding requires much more than a list of the objects present and their locations. To understand a scene, plan action on it and predict what will happen next, we must extract the relationships between objects (such as support and attachment), their physical properties (such as mass and material) and the forces acting on them. One view is that this is done with a "mental physics engine" that represents this information and runs forward simulations to predict what happens next.
This keynote reviews fMRI evidence that regions in the parietal and frontal lobes, but not the ventral visual pathway, behave as expected of such an engine: they respond more strongly when deciding about physical than visual properties and when viewing physical rather than social stimuli, and they contain scenario-invariant information about object mass inferred from motion trajectories, the stability of object configurations and whether two objects are in contact. Most tellingly, predicted collision events can be decoded from perceived collision events, as expected if these regions run forward simulations. The talk also considers how far the system is engaged by fluid "stuff" as well as rigid "things", and in some circumstances by language, and argues that these findings, together with the poor performance of deep network models on many intuitive physics tasks, provide preliminary evidence for a physics engine in human parietal and frontal cortex.