ABSTRACT
The violation-of-expectation paradigm investigates infants’ physical knowledge by exploiting their tendency to look longer at events that are surprising, unexpected, or physically impossible. The current simulation study examines the role of prediction as a fundamental component of infants’ expectations in physical-knowledge studies. A recurrent network is presented with a computer-animated version of Baillargeon’s “car study” (1986; Baillargeon & DeVos, 1991), in which a car rolls down a ramp and behind a screen. After learning to predict the outcome of a training event, the model is then tested on possible and impossible events from the same study. During testing, the model successfully predicts only superficial features of the test events. These results are used to argue for the necessity of prior physical knowledge, and perhaps also a built-in capacity for mental representation, in order for a prediction system to work.
