A hierarchical Bayesian model of conceptual knowledge transfer


Agents generalise abstract conceptual knowledge across different contexts. For example, an individual negotiating a new computer program will draw upon experience with similar programs, such as how to use a drop-down menu. What are the rules governing such knowledge transfer? Here we offer a formal Bayesian account of generalisation, in which observers update a hierarchical model that incorporates knowledge about the statistical moments of the distribution from which information is drawn. We use this model to predict performance on a foraging task that involved hunting for hidden rewards in a virtual two-dimensional grid environment. In this task, contextual cues signalled not only the likely reward location (bivariate mean), but also the pattern (bivariate dispersion). Observers optimally integrated noisy cues about the probable reward location with information from these cues. This model and data offer a formal account of how humans learn abstract conceptual information.

Back to Table of Contents