Abstract:

Infants as young as 6 months of age appear to be sensitive to whether observed agents behave in an efficient manner when acting towards a goal. This has led some researchers to posit an innate "rationality principle" which guides infants' reasoning about actions and facilitates online prediction of others' behavior. However, infants also seem to be sensitive to frequency information and agency-related perceptual cues that implicate more general, statistical-learning processes in which the structure of actions is gradually represented by generalizing over many similar events. I will present an instantiation of statistical learning--a Parallel Distributed Processing (PDP) model-- that was trained to predict curvilinear agent motion in the presence of goals and constraints in an artificial environment. Simulation results indicate that a) looking patterns of infants from the classic experimental paradigm can be generated as a function of experience, without a need for a rationality principle and with less strict assumptions than existing models; b) our model also shows a similar, age-varying pattern of dependence on perceptual and motion cues observed in infants between 6- and 12-months of age; and c) perceptual prediction, rather than being a consequence of the rationality principle, may in fact be a primary driver of infants' early learning about actions.