Recursive, rich, and rational: characterizing learners' inferences in pedagogical learning throughout development

From a young age, humans learn about the world from other people. Past computational models of pedagogy highlight the reciprocal reasoning processes that support learning from others. But social learning can be messy, often violating the assumptions of these models. How does learning from pedagogy remain robust in the face of non-ideal circumstances? Through empirical studies with adults and children, this dissertation investigates three predictions that fall out of pedagogical models when their default assumptions are violated. These processes are further formalized using Bayesian models of cognition. First, how do learners evaluate the quality of potential informants to avoid imperfect teaching? I propose a computational model of teacher evaluation that considers the utility of a teacher’s demonstrations given their prior knowledge. Adults’ ratings of teachers’ quality aligned with the model’s predictions; however, these abilities were tenuous in preschoolers. By second grade, while children considered demonstrations’ utility, only those children who spontaneously invoked the teacher’s knowledge as a relevant causal explanatory variable considered this factor in their ratings. Second, how might learners’ expectations about informants’ teaching styles shape learning from new demonstrations? I suggest that inferences about (1) the amount of information to learn, and (2) the importance of demonstrations, should be rationally informed by expected teaching style. Adult behavioral data supported these predictions, and modeling of these results explains how these inferences could be made. However, model predictions were not reflected in an exploratory play task with preschoolers; instead, only age and socioeconomic status predicted patterns in children’s play. Finally, if a learner knows that a teacher has inaccurately represented their prior knowledge, how might future learning from that teacher be calibrated? A causal model predicts that if a teacher underestimates a learner’s prior knowledge, that learner will seek out additional challenge in a subsequent task; but this challenge-seeking might be curbed if a teacher instead overestimates that learner’s prior knowledge. Results from children and adults supported these predictions. This work informs our understanding of how learning is shaped by the people around us throughout our lives, and extends our models of social cognitive development to diverse and novel contexts.Ph.D.Includes bibliographical reference