Learning Children’s Conceptual Spaces using Deep Metric Learning.

Children learn to represent the world around them in meaningful categories that allow them to generalize past experiences. Understanding how these categorical representations develop is fundamental to cognitive science. However, capturing the structure of human conceptual knowledge is a challenging experimental task. The most prominent approach, Multidimensional Scaling (MDS), usually requires participants to produce many similarity judgments, leading to long experiments. Moreover, the representations found by MDS are limited to the fixed set of experimental stimuli and have to be reconstructed for every new item. In contrast, we present a more flexible machine-learning method that can generalize to novel stimuli. This method uses a child-friendly task that allows researchers to uncover the development of categories with fewer participant judgments. We evaluate our approach on simulated data and find that it can accurately reveal representations even when trained on data generated by groups that categorize differently. We then analyze data from the World Color Survey and find that we can recover language-specific color organization when aggregating languages that only share the same number of basic color terms. Finally, we use the method in a developmental experiment and find age-dependent differences in how complex fruit stimuli are organized. These differences were consistent with participants' reasoning and additional experimental measures. Our results suggest that our approach is applicable in psychological tasks and opens the possibility of examining children's developing psychological spaces in new detail.