Abstract
Engaging with the world requires a model of its physical structure and dynamics – how objects rest on and support each other, how much force would be required to move them, and how they behave when they fall, roll, or collide. Humans demonstrate remarkable ability to infer the physical properties of objects and predict physical events in dynamic scenes, yet little is known about the neural representations underlying intuitive physical judgments. Recent behavioral and computational studies of human physical scene understanding suggest that people's judgments can be modeled as probabilistic simulations of a mental physics engine akin to 3D physics engines used in computer simulations and video games. Physics engines share a common structure: enduring properties of objects such as mass and friction serve as inputs to models of world dynamics. We ask whether such a physics engine exists in the brain, and begin by searching for neural representations of fundamental physical variables that define objects. Here, using event-related fMRI and multivariate pattern classification techniques, we tested whether candidate functional regions of interest (fROIs) for a neural physics engine (Fischer et al., 2016) represent object mass. We obtained significant mass decoding in 12 out of 12 subjects from multivoxel activity within these fROIs while adult subjects watched videos of objects interacting in various physical scenarios: splashing into water, being blown by a hairdryer, and falling onto a soft surface. Critically, these mass representations generalize across scenarios, and were decoded during both a mass judgment task and an orthogonal task (a color judgment on the same stimuli). These results suggest that candidate fROIs for a neural physics engine represent situation-invariant physical information that may serve as input to a generalized engine for physical simulation. Ongoing work investigates what other physical information is represented, and the generality and automaticity of these representations.
Meeting abstract presented at VSS 2018