To what extent does the brain simulate the external world?
Kevin Smith, Massachusetts Institute of Technology
Todd Gureckis, New York University
Marcelo Mattar, University of California, San Diego
Kelsey Allen, DeepMind
Fred Callaway, Princeton
Ernest Davis, New York University
Judith Fan, University of California, San Diego
Tom Griffiths, Princeton
Jessica Hamrick, DeepMind
Wei Ji Ma, New York University
Kimberly Stachenfeld, DeepMind
Joshua Tenenbaum, Massachusetts Institute of Technology
Tomer Ullman, Harvard
Our ability to imagine future and past states of the world helps us reason, infer and plan. We can predict what will happen when dishes are too precariously stacked, which businesses we might pass on a drive to a friend’s house, and how our lives and careers might unfold. This ability seems critical to our ability to act in the world, allowing us to avoid negative consequences, and to create desired outcomes. However, the computational basis of this ability remains poorly understood. One popular idea is that we simulate possible pasts and futures using some explicit model of the world. Such ideas have been proposed in many areas of cognitive neuroscience including decision making (Klein & Crandall, 2018), planning (van Opheusden et al., 2021), memory (Schacter, 2022), and intuitive reasoning about the social and physical world (Allen, Smith, & Tenenbaum, 2020; Baker, Saxe, & Tenenbaum, 2009; Battaglia, Hamrick, & Tenenbaum, 2013). While theoretically appealing, the simulation hypothesis is not always necessary to account for the observed data, nor is it rigorously compared to alternative approaches. The purpose of this GAC is to explore the degree to which the mind and brain uses simulation, what actually constitutes simulation-based cognition, and how the brain might simplify representations and processes to more efficiently reason about and act in the world. This GAC team takes a broad, interdisciplinary perspective on this issue, and draws on research in psychology, AI, and neuroscience, spanning a number of domains in which simulation has been theorized to be used. The ultimate output of the GAC will be (i) a description of the various ways that simulation is used throughout cognitive computational neuroscience, including common definitions of simulations across domains and the types of empirical evidence both in favor of and against the use of simulation; (ii) a theoretical framework to understand how simulation is constrained by the brain and how the brain overcomes those constraints; and (iii) a set of experiments to test the claims of any proposed frameworks.