[PS-3.17]The Nested Timescales of Statistical Learning: A Dynamic Field Model of The Shape Bias

Samuelson, L. 1, 3 , Perone, S. . 2, 3 & Spencer, J. 1, 3

1 University of East Anglia
2 Washington State University
3 The DeLTA Center

Increasingly, theoretical and modelling work has highlighted the need to understand how statistical learning is integrated with perceptual, memory, and attentional processes both neurally and in behaviour. To this end, we present a neurally-grounded, process model of the shape bias. The shape bias is the tendency of young children to generalize novel nouns for novel solid objects according to similarity in shape. The extant literature shows that this bias is the product of accumulating statistics across many individual moments of word learning and word use. Thus, it lives at timescales of both real-time behaviour-generalizing a novel noun in the context of an ongoing interaction-and of the developing vocabulary. Furthermore, it is developmentally related to changes in object perception, attention, and memory. Five sets of simulations show that our Dynamic Field model captures the emergence of the bias from the statistics of the vocabulary, subsequent accelerations in vocabulary development, and changes in the bias related to task context and individual differences in children?s vocabularies. Because the DF model embeds these phenomena in a process model of memory and decision making, it grounds statistical learning in these core cognitive processes and demonstrates the nested timescales of statistical learning.