Computational Models of Intuitive Physics

People have a powerful “physical intelligence” – an ability to infer physical properties of objects and predict future states in complex, dynamic scenes – which they use to interpret their surroundings, plan safe and effective actions, build and understand devices and machines, and communicate efficiently. For instance, you can choose where to place your coffee to prevent it from spilling, arrange books in a stable stack, judge the relative weights of objects after watching them collide, and construct systems of levers and pulleys to manipulate heavy objects. These behaviors suggest that the mind relies on a sophisticated physical reasoning system, and for decades cognitive scientists have been interested in the content of this knowledge, how it is used and how it is acquired. In the last few years, there has been exciting progress in answering these questions in formal computational terms, with the maturation of several different traditions of cognitive modeling that have independently come to take intuitive physics as a central object of study. The goals of this symposium are to: 1) highlight these recent computational developments, focusing chiefly on qualitative reasoning (QR) models and Bayesian perceptual and cognitive models; 2) begin a dialog between leading proponents of these different approaches, discussing a number of dimensions along which the approaches appear to differ and working towards bridging those differences; 3) enrich these models with perspectives from empirical work in cognitive science. Background. The research to be discussed builds on several decades of prior work from multiple traditions in cognitive science. Cognitive psychologists since the 1970s have studied the role that human intuitive physics plays in development, perception, education, and reasoning. Behavioral research with adults focused on identifying errors and biases in people's general understanding and theories about physical rules (McCloskey, 1983), as well as psychophysical studies of how sensory cues drive specific judgments in dynamic displays (Todd & Warren, 1982). Early and ongoing developmental work has identified milestones in cognitive sensitivity and expectations about core physical principles (Baillargeon, 2007). Though these efforts have made significant progress, they did not frame their results as computational models with sufficient clarity and power to explain people's physical reasoning in complex and varied scenes. Crucial computational progress has come from the fields of human and computer vision, artificial intelligence (AI), and machine learning. Human and machine vision researchers have recently developed computational models of natural scene understanding (Oliva & Torralba, 2007), but their focus has been on knowledge about the geometry and semantics of scene layouts, not the role of physical constraints and how physical properties are represented and exploited for prediction, reasoning and planning. AI researchers have been developing frameworks for qualitative reasoning (QR) and applying them to physical domains for over 30 years, and these approaches have now matured to the point that they can both solve challenging real-world inference problems and engage directly with behavioral experiments, giving state-of-the-art accounts of people’s intuitive reasoning in a wide range of science and engineering domains (Forbus, 2011). The framework of Bayesian reasoning in probabilistic generative models has revolutionized AI and machine learning, and in the last decade has also come to provide a lingua franca for sophisticated reverse-engineering models of human perception, action and cognition (Chater et al, 2006; Tenenbaum et al, 2011). But only in the last few years have Bayesian models been applied to challenging physical reasoning problems, and been shown to give strong quantitative accounts of human physical judgments (Sanborn et al, 2009; Hamrick et al, 2011). This symposium brings together leading researchers modeling intuitive physics from the QR, Bayesian cognition and perceptual modeling traditions, to discuss highlights of recent models and points of contact and contrast between different modeling approaches. The talks and discussion will explore several axes in the space of possible models, including the following: rational reverse-engineering vs. descriptive or heuristic accounts; qualitative vs. quantitative reasoning; probabilistic vs. deterministic inference; lowerlevel perceptual vs. higher-level cognitive inferences; implicit vs. explicit reasoning; analog simulation vs. symbolic rule-based representations; the role of memory-, experienceand learning-dependent reasoning; the role of