Development of Causal Theories

Tom Griffiths
University of California, Berkeley (UC Berkeley)

Causal theories provide constraints that are necessary for rapid learning of causal relationships, but present a new set of challenges for a learner: how can the learner learn the theories themselves? I will discuss how we can approach this question using the tools of hierarchical Bayesian models, and describe some recent experiments exploring learning of causal theories in both children and adults.
These experiments show that both children and adults are capable of learning the probabilities with which causal relationships exist and the form of those relationships, producing results that are consistent with Bayesian inference being applied not just at the level of hypotheses about whether a causal relationship exists, but also at the higher level of causal theories.

Presentation (PDF File)

Back to Long Programs