Constraints on Bayesian Explanation

Introduction The hypothesis that human cognition may be well characterized as a set of Bayesian computations has been the topic of considerable debate over the last two decades. Recently, critics have argued that this hypothesis is either unlikely to be true or otherwise too unconstrained to be particularly useful for explaining cognition (e.g., Bowers & Davis, 2012), whereas proponents have defended their position by stating that the Bayesian perspective has been misunderstood, is not necessarily in conflict with other perspectives on cognition, and can still be explanatorily useful as a framework for cognitive science even if underconstrained in many ways (e.g., Griffiths, Chater, Norris, & Pouget, 2012). Our position in this debate is that both sides of this debate may be right as well as wrong: Proponents may be right that the Bayesian perspective has something uniquely useful to bring to cognitive science (and then the critics are wrong in their denial of this); yet, the critics may be right that cognitive theories are explanatorily useful only if properly constrained (and then proponents are wrong in their denial of this). With this perspective in mind, we wish to move the debate forward in a constructive way by bringing in new perspectives and proposing novel constraints that can be exploited for purposes of improving the explanatory values and virtues of Bayesian explanations of cognition. Specifically, with this symposium we aim to focus on how constraints on Bayesian explanations can be exploited in ways that are yet underrepresented and underexplored. The symposium brings together researchers from various disciplines, contributing a variety of perspectives on how Bayesian explanations can be fruitfully constrained, drawing on theories, analyses, and results from philosophy of science, cognitive neuroscience, information theory, machine learning, and theoretical computer science.