Intuitive Theories
By Tobias Gerstenberg, Josh Tenenbaum. 2020
One of the hallmarks of human intelligence is its flexibility. While computers have achieved better-than-human performance in various games, such as Chess, Jeopardy, or Go, there is no single algorithm yet that works in all of these cases. Humans, however, can excel at all of these games, and many other tasks. We believe that bridging the gap between human and machine intelligence requires two key insights from cognitive science: (i) human knowledge is organized in terms of richly structured intuitive theories, and (ii) many cognitive processes can be understood as causal inferences operating over these structures. In this chapter, we first explain what intuitive theories are, how we can model them as probabilistic, generative programs, and how intuitive theories support various cognitive functions such as prediction, counterfactual reasoning, and explanation. We focus on two domains of knowledge: people’s intuitive understanding of physics, and their intuitive understanding of psychology. We show how causal judgments can be modeled as counterfactual contrasts operating over an intuitive theory of physics, and how explanations of an agent’s behavior are grounded in a rational planning model that is inverted to infer the agent’s beliefs, desires, and abilities. We conclude by highlighting some of the challenges that the intuitive theories framework faces, such as understanding how intuitive theories are learned and developed.
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
- Bernhard Schölkopf
- Soujanya Poria
- Jiafei Duan
- Giambattista Parascandolo
- Samson Yu
- Niki Kilbertus
- Percy Liang
- Cheston Tan
Citations in the corpus
- Generalization in anti-causal learning (2018, background): “It has been conjectured that humans internally build powerful causal models for simulations and approximate physical mechanisms through intuitive theories (Krynski and Tenenbaum, 2007; Waldmann et al., 2006; Hamrick et al., 2011; Battaglia et al., 2013; Gerstenberg and Tenenbaum, 2017).”
- Explanations in the wild. (2021, unknown): “…(Prasada et al., 2013), inherent essences (Cimpian & Salomon, 2014; Cimpian & Steinberg, 2014), category judgments (Rehder & Hastie, 2001), and even high-level epistemic frameworks such as intuitive theories (Gerstenberg & Tenenbaum, 2017) or models of understanding (Lombrozo & Wilkenfeld, 2019).”
- PIP: Physical Interaction Prediction via Mental Imagery with Span Selection (2021, methodology): “We use noisy and approximate physical simulations to account for property, perceptual, dynamic and even collision uncertainties [15, 22, 23, 24, 25, 26, 17].”