Learning Physical Parameters From Dynamic Scenes

Abstract

Humans acquire their most basic physical concepts early in development, and continue to enrich and expand their intuitive physics throughout life as they are exposed to more and varied dynamical environments. We introduce a hierarchical Bayesian framework to explain how people can learn physical parameters at multiple levels. In contrast to previous Bayesian models of theory acquisition (Tenenbaum et al., 2011), we work with more expressive probabilistic program representations suitable for learning the forces and properties that govern how objects interact in dynamic scenes unfolding over time. We compare our model to human learners on a challenging task of estimating multiple physical parameters in novel microworlds given short movies. This task requires people to reason simultaneously about multiple interacting physical laws and properties. People are generally able to learn in this setting and are consistent in their judgments. Yet they also make systematic errors indicative of the approximations people might make in solving this computationally demanding problem with limited computational resources. We propose two approximations that complement the top-down Bayesian approach. One approximation model relies on a more bottom-up feature-based inference scheme. The second approximation combines the strengths of the bottom-up and top-down approaches, by taking the feature-based inference as its point of departure for a search in physical-parameter space.

A counterfactual simulation model of causal judgments for physical events.

“…The CSM is a model of causal judgment and not of causal learning Goodman, Ullman, & Tenenbaum, 2011;Kemp, Goodman, & Tenenbaum, 2010;Lake, Ullman, Tenenbaum, & Gershman, 2016;Tenenbaum, Kemp, Griffiths, & Goodman, 2011;Ullman, Goodman, & Tenenbaum, 2012;Wellman & Gelman, 1992). The CSM assumes that people already possess a causal model that incorporates the relevant domain knowledge for simulating different counterfactuals (see Bear et al, 2020;Ullman, Stuhlmüller, Goodman, & Tenenbaum, 2018;Yi* et al, 2020, for work on how physical models may be learned). However, we don't assume that people's causal model is perfectly accurate, and we capture this uncertainty by introducing noise into the physical simulation model as described below (Smith et al, submitted;Ullman, Spelke, Battaglia, & Tenenbaum, 2017).…”

“…The CSM assumes that people already have access to a generative model of the domain. Recent work has looked into how it may be possible to learn such a causal model from observing and interacting with the world (Baradel, Neverova, Mille, Mori, & Wolf, 2019;Battaglia et al, 2018;Bramley, Gerstenberg, Tenenbaum, & Gureckis, 2018;Ullman et al, 2018;Yi* et al, 2020). If a person already has an accurate model of the world, what role, if any, does the ability to make causal judgments play?…”

Limits on simulation approaches in intuitive physics

“…In many cases, maintaining consistency is critical if the simulator is to give results that are correct, or even physically plausible. For example, Ullman, Stuhlmüller, Goodman, and Tenenbaum (2018) used a PMST model to investigate people's ability to infer the properties of multiple pucks interacting on a two-dimensional plane, where the participants saw clips of the pucks interacting and then were asked to estimate properties such as the masses of the pucks and the roughness of certain colored patches of the two-dimensional plane. Based on behavioral data, the authors conclude that features might be learned by running an intuitive physics simulator with different parameter values many times, and accepting the parameters which generate simulations most similar to what was observed.…”

“…To test this, we created simple 2-D worlds, in which a small number of objects interacted with one another in the presence of gravity. We chose this design to remain similar to the materials used in some previous PMST work (Smith et al, 2017;Ullman et al, 2018).…”

Physically Implied Surfaces

“…More generally, attention and cognition are especially tuned to physical interactions in the world. For example, infants are sensitive to causality in displays of collisions (Leslie & Keeble, 1987), and adults readily infer properties of objects involved in physical interactions (Hamrick, Battaglia, Griffiths, & Tenenbaum, 2016; Todd & Warren, 1982; Ullman, Stuhlmüller, Goodman, & Tenenbaum, 2018), represent the future states of moving and colliding objects (Gerstenberg, Peterson, Goodman, Lagnado, & Tenenbaum, 2017; see also Guan & Firestone, 2020; Hubbard & Bharucha, 1988; Peng, Ichien, & Lu, 2020), and even form impressions of causal relations mediated by unseen, force-transmitting connecting elements (as in classic “pulling” stimuli; Michotte, 1963; White & Milne, 1997; see also Scholl & Nakayama, 2004).…”