Abstract
Drawing is a powerful tool for communicating concepts in visual form — a few well-placed strokes can convey the identity of a person, object, or scene. Prior work has found that deep neural network models of the ventral stream trained purely on photographs can also recognize drawings by nonexpert adults, reflecting concordance in abstract representations of object categories in drawings and photos at higher layers in these models (Fan, Yamins, & Turk-Browne, 2015). How do ordinary people become so effective at producing recognizable drawings? Here we examine the trajectory of this learning during childhood. Children (N = 41, M = 6.9 years, range 4-10 years) participated in an iPad-based drawing game where they were prompted with a verbal cue to draw one of sixteen familiar objects (e.g., "Can you draw a cup?"). Children drew each object category for 30 seconds, after which they were prompted to either make another drawing or to stop drawing altogether. Afterwards, a group of naive adults (N = 14) guessed the identity of each drawn object (286 drawings). A generalized logistic mixed-effect model revealed that the recognizability of drawings increased reliably with age (b = 0.96, SE = 0.17, Z = 5.5), accounting for variation across object categories and individual children (% drawings recognized; chance = 4.8%; M4yrs = 14%, M5yrs = 45%, M6yrs = 70%, M7yrs = 72%, M8yrs = 66%, M9yrs = 76%, M10yrs = 85%). Further, this relationship persisted when controlling for several low-level covariates — the amount of time spent drawing, the number of strokes, and total ink used. These results suggest that the capacity to quickly produce graphical representations that communicate object category information is highly developed by middle childhood. More broadly, these findings point to visual production tasks as a promising avenue for examining the development of object category representations.
Meeting abstract presented at VSS 2018