GATSBY COMPUTATIONAL NEUROSCIENCE UNIT

Blicket detectors and theory-based Bayesian inference

Accounts of human causal induction have traditionally tended to emphasize the importance of either causal mechanism knowledge or covariational evidence. I will argue that in order to explain how people can infer causal relationships from small amounts of data, we need to consider how these two sources of information are combined. To this end, I will outline an account of causal induction that uses "theory-based" Bayesian inference, where a causal theory generates a hypothesis space of possible causal structures that are evaluated using Bayes' rule. This account will be illustrated through a series of experiments using the blicket detector, a device that has been used to study causal learning in both children and adults. These experiments were designed to test the two key predictions of the theory-based Bayesian account: that learners should be sensitive to the prior probability that a causal relationship exists, and that the conclusions drawn by learners will be affected by whether the mechanism by which the detector operates is probabilistic or deterministic. This is joint work with Josh Tenenbaum, David Sobel, and Alison Gopnik.