Infants generate structured learning environments during curiosity-driven category exploration

Ke. H., Westermann, G., & Twomey, K. E. (forthcoming; 2018, July). Infants generate structured learning environments during curiosity-driven category exploration. Poster to be presented at Expanding the Field 2018: Multidisciplinary Developmental Dynamics Workshop, University of East Anglia, Norwich, UK

Decades of developmental research offer a rich picture of the factors implicated in the fundamental skill of early category learning. However, with a few exceptions, these studies typically control stimulus order and timing a priori. Outside the lab, in contrast, infants explore their learning environment freely. Surprisingly, given the ubiquity of this curiosity-driven learning in early development, it is yet to be fully characterised. However, recent computational work (Twomey & Westermann, 2017) predicts that systematicity in curiosity-driven exploration should unfold dynamically as infants move from stimulus to stimulus, generating an intermediate level of task complexity. The current study tests this prediction in a free exploration task.

To quantify complexity we developed three categories of 3D printed toy objects, each consisting of five exemplars varying along a continuum. For example, the pyramids category consisted of a triangular pyramid, then a four-sided pyramid, then a five-sided pyramid, and so on. Twelve-, 18- and 24-month-old children (N = 54) then took part in a shape priming task (Twomey, Malem & Westermann, 2016). Children saw six trials (two per category), during which they played with a prime object from one end of the continuum for 15 seconds, then the remaining four objects 30 seconds.

To test the hypothesis that infants explore systematically, we coded for first touches and transitions between objects. Transitions between objects were a proxy for perceptual distance; for example, a transition from the second object in a category to the third constituted a distance of 1, and a transition from the second object to the fifth was a distance of 3.

First touches were non-random (12mo: χ2(3) = 12.09, p = .001; 18mo: χ2(3) = 22.84, p < .001; 24mo: χ2(3) = 8.05, p = .045); specifically, infants initially systematically chose the most distant object, irrespective of prime. Next, we submitted exploratory sequences to a hierarchical cluster analysis to determine whether infants’ dynamic exploration was also structured. These analyses detected clusters in all age groups’ exploratory sequences (12mos: 3; 18mos: 3; 24mos: 2). To characterize these clusters, we coded each sequence by prime, mean transition distance, sequence length, and entropy. We then used these indices to predict cluster membership in a series of multinomial regressions. Across ages, mean distance cluster membership (all ps < .0051); specifically, infants systematically generated sequences of intermediate distances. Cluster membership was also predicted by sequence length (all ps < .00023). 

In line with Twomey & Westermann (2017), infants generated systematic exploratory sequences when engaged in curiosity-driven object exploration. Specifically, the unfolding of these exploratory sequences in time indicates that infants prefer intermediate complexity (cf., Kidd, PIantadosi & Aslin, 2012). Overall, this work demonstrates that infants are neither passive learners nor random explorers; rather, they structure their own learning environment. Thus, we highlight the importance of taking seriously this dynamic learner-environment interaction for a clear understanding of development.