Technical Program

Paper Detail

Paper: PS-1A.25
Session: Poster Session 1A
Location: Symphony/Overture
Session Time: Thursday, September 6, 16:30 - 18:30
Presentation Time:Thursday, September 6, 16:30 - 18:30
Presentation: Poster
Paper Title: Modeling the Intuitive Physics of Stability Judgments Using Deep Hierarchical Convolutional Neural Networks
Manuscript:  Click here to view manuscript
Authors: Colin Conwell, George Alvarez, Harvard University, United States
Abstract: To gauge whether a tower of objects will fall, are visual heuristics sufficient? In this study, we explore the potential of pattern recognition as a viable model of intuitive physical inference in a task that requires observers to estimate the stability of stacked building blocks, comparing the performance of a deep feedforward convolutional neural network to human performance using psychophysics. In analyzing human and machine behavior alike, we identify a pair of image-based visual features that strongly predict both human and machine performance and differ only in the summary statistic used to compute them. Our results suggest that a system trained only to recognize patterns in visual input and given no explicit physical knowledge (e.g. mass, gravity, friction or elasticity) is nonetheless capable of approximating human judgments in a paradigmatic intuitive physics task.