Inferring mass in complex physical scenes via probabilistic simulation
Originally presented at MathPsych 2013 in Potsdam, Germany.
How do people learn about underlying physical parameters, such as mass and friction, from the interactions of objects in complex, dynamical scenes? Discovering the values of such properties is a difficult task: the parameters cannot be observed directly, and their influence on the sequence of dynamics is often complicated and difficult to resolve unambiguously. Yet, people can do this. We showed participants towers of blocks stacked in complex configurations, and asked them to predict whether each tower would fall. After giving them the correct answer, we asked a further question: which blocks were heavier? With the correct information about stability, participants rapidly learned the blocks' relative masses. We propose that this learning can be viewed as probabilistic inference in a generative model that approximates Newtonian rigid-body dynamics. To express this hypothesis, we built a model that uses physical simulation and Monte Carlo sampling to predict what will happen and then to update its beliefs based on the divergence of its predictions from reality. Participants' judgments were qualitatively consistent with those of this physics-aware model observer, but also deviated in key ways that may be explained by information and resource limitations. This is an important step in understanding how people perceive and reason about the physical nature of their environment, and provides a working framework for modeling and testing people's inferences about unobserved properties in complex, real-world scenes.
Slides created using the IPython Notebook (http://ipython.org/) and reveal.js (http://lab.hakim.se/reveal-js/) and can be found on my website (http://www.jesshamrick.com/)
Presentation created using Camtasia for Mac (http://www.techsmith.com/camtasia.html)