Inferring Mental Representations with Bayesian Models of Cognition

Bayesian models of cognition (BMCs) have typically been used in a normative manner to characterize the rational solutions to inductive inference problems in the environment. I discuss the potential of re-framing these models as generative processes for human behavior. In this approach, computational-level assumptions about the environment are reinterpreted as psychological assumptions about cognitive processes and mental representations. Unlike with BMCs, these re-framed models define the probabilistic relationship between model assumptions and human behavior, and can therefore be used to draw inferences about cognitive processes and mental representations from human data. Furthermore, standard model selection techniques can be used to evaluate and compare these models against competing Bayesian and non-Bayesian models. I present two studies---the first involving predictions about everyday events and the second involving a reconstructive memory task---in which BMCs are re-framed as generative process for human behavior in order to infer the subjective knowledge of participants based on their response data. In these studies, we find that when compared to models assuming people's knowledge is veridical with environmental statistics, models using inferred subjective representations---including individual differences---provide better generalization to unobserved behavioral data. A third study demonstrates the limitations of BMCs when attempting to model the perspectives of multiple individuals in a social cognition task, and provides motivation for the new approaches developed in this dissertation. Finally, we consider theoretical implications of this approach in relation to standard Bayesian models of cognition.