Picture completion reveals developmental change in representational drawing ability: An analysis using a convolutional neural network

Abstract

Drawings of children may provide unique insights into their cognition. Previous research showed that children's ability to draw objects distinctively develops with increasing age. In recent studies, convolutional neural networks have been used as a diagnostic tool to show how the representational ability of children develops. These studies have focused on top-down task modifications by asking a child to draw specific objects. Object representations, however, are influenced by bottom-up visual perception as well as by top-down intentions. Understanding how these processing pathways are integrated and how this integration changes with development is still an open question. In this paper, we investigate how bottom-up modifications of the task affect the representational drawing ability of children. We designed a set of incomplete stimuli and asked children between two and eight years to draw on them without specific task instructions. We found that the higher layers of a deep convolutional neural network pretrained for image classification on objects and scenes well differentiated between different drawing styles (e.g. scribbling vs. meaningful completion), and that older children's drawings were more similar to adult drawings. By analyzing representations of different age groups, we found that older children adapted to variations in the presented stimuli in a more similar way to adults than younger children. Therefore, not only a top-down but also a bottom-up modification of stimuli in a drawing task can reveal differences in how children at different ages represent different concepts. This task design opens up the possibility to investigate representational changes independently of language ability, for example, in children with developmental disorders.