Second-Order Correlation Learning of Dynamic Stimuli: Evidence from Infants and Computational Modeling.

Abstract

We present two habituation experiments that examined 20- and 26-month-olds' ability to engage in second-order correlation learning for static and dynamic features, whereby learned associations between two pairs of features (e.g., P and Q, P and R) are generalized to the features that were not presented together (e.g., Q and R). We also present results from an associative learning mechanism that was implemented as an autoencoder parallel distributed processing (PDP) network in which second-order correlation learning is shown to be an emergent property of the dynamics of the network. The experiments and simulation demonstrate that 20- and 26-month-olds as well as neural networks are capable of second-order correlation learning in a category context for internal features of dynamic objects. However, the model predicts-and Experiment 3 demonstrates-that 20- and 26-month-olds are unable to encode second-order correlations in a noncategory context for dynamic objects with internal features. It is proposed that the ability to learn second-order correlations represents a powerful but as yet unexplored process for generalization in the first years of life.