Carvalho, P. & Goldstone, R.
Indiana University, Bloomington, IN, USA
Learning happens over time. Acquiring new information is the result of a dynamic process of selecting, storing and retrieving relevant information while forgetting irrelevant information available at each moment. In this paper we look at how the statistics of the learning situation affect learning by changing what is encoded. We compare two study sequences with differing category transition probabilities: One in which categories alternate very frequently and one in which the same categories alternate infrequently. The results, across different types of categories and different types of learning tasks, show that the probability of category alternation changes what people attend to during learning and consequently what is encoded and learned. During study, learners' attention is biased towards the properties emphasized by the transition probabilities of the study sequence (differences between categories in high alternation sequences and similarities within categories for low alternation sequences). What information is encoded predicts the effectiveness of learning and depends on the sequential statistics of the learning situation. These results are captured by an exemplar model that takes into account the sequence of exemplars during learning by changing the likelihood of attending to and encoding different object properties depending on sequential similarities.