Extending and testing the bayesian theory of generalization

Abstract

We introduce a tractable family of Bayesian generalization functions. The family extends the basic model proposed by Tenenbaum and Griffiths (2001), allowing richer variation in sampling assumptions and prior beliefs. We derive analytic expressions for these generalization functions, and provide an explicit model for experimental data. We then present an experiment that tests the basic model predictions within the core domain of the theory, namely tasks that require people to make inductive judgments about whether some property holds for novel items. Analysis of the results illustrates the importance of describing variations in people’s prior beliefs and assumptions about how items are sampled and of having an explicit model for the entire task.