Figure 5.

Schematic of Hierarchical Bayesian Model (HBM), as applied to transitive inference task. Overview of the generative HBM of Kemp and Tenenbaum (2008): the model is specified at two levels: a high “structure” level that specifies the type of structure that best explains data—in this case a hierarchy of kanji characters (e.g., used in experiment by Greene et al., 2006)—a generative process based on graph grammars is used to create a library of possible structures. A lower “instance” level specifies the exact version of the structure that is most likely given the data—in this case A>B>C>D>E. The model simultaneously finds the best structure and instance that likely accounts for the data. Upward arrows indicate the generative nature of the HBM.