Developmental changes in drawing ability: Comparing adult ratings and convolutional neural network analysis

• In Event: Poster Session 14
  In Poster Session: PS 14 Topic - Cognitive Processes

Abstract

Children’s drawing ability develops gradually with increasing age and reflects the maturation of the underlying cognitive processes. One mechanism that might explain the underlying cognitive processes of drawing is predictive coding: by integrating top-down predictions with bottom-up sensory perceptions, predictions can be made about how, for example, a partially shown stimulus should be completed. Adults are thought to optimally integrate top-down and bottom-up information [Friston & Kiebel 2009], however, it is an open question how this process changes over the course of development.

One way to study this is by looking at how children complete partial drawings where crucial features are missing, a task that was first designed by [Saito et al. 2014]. However, one shortcoming of previous studies is that they were based on small sample sizes and subjective evaluations of the drawings. Here, we extend the drawing task of [Saito et al. 2014] to investigate how children’s drawings change over the development. The task was to complete partial drawings of objects such as a face or a house. Either the object’s outline, the inner features or the inner features in a scrambled configuration were presented to investigate how modifications in the bottom-up stimuli affected the children’s drawings. We collected more than 600 drawings from 103 children, (2-8 years, 61 male).

To overcome limitations of previous studies, we analyse the data in two distinct ways: (1) by adult ratings obtained via crowdsourcing and (2) using quantitative features, namely, the activation patterns of a convolutional neural network (CNN) based on a method proposed recently in [Long et al. 2018].

The results demonstrate the developmental change of children’s drawing ability. Analysis (1) confirmed our hypotheses that children completed more (chi_square(1)=446.53, p<.001) and scribbled less (chi_square(1)=467.69, p<.001) with increasing age (see Figure 1).

For analysis (2), we divided the children into four age groups, equally distributing the amount of drawings contained in each group. The activations of the highest layer of the CNN in response to each drawn picture are computed. From the distance between the features when the child drew on one stimulus vs. on another stimulus, the representational dissimilarity matrix can be constructed [Long et al. 2018]. Figure 2 displays this matrix, averaged for the different age groups, and additionally of drawings obtained from four adults. It can be observed that young children drew similarly on all presented stimuli whereas adults and older children adapted their drawings to differences in the presented stimuli. This structure gradually develops with increasing age, confirming the developmental change of drawing.

Our findings confirm that the CNN analysis (a) well reflects results obtained from traditional analyses based on an adult rating study, and (b) demonstrate that the developmental change can be measured using bottom-up modifications of the task stimuli. Unlike the task in [Long et al. 2018], this task design does not require linguistic instructions to the children, opening up the possibility to investigate drawing in young children and in children with linguistic impairments such as developmental disorders.

Authors

• Anja Philippsen, The University of Tokyo
  Presenting Author

• Sho Tsuji, The University of Tokyo
  Non-Presenting Author

• Yukie Nagai, The University of Tokyo
  Non-Presenting Author