Abstract
In the study of language, production has long been viewed as essential for understanding comprehension. By contrast, vision science has centered almost exclusively on comprehension (recognition) and has largely neglected production. Here we investigate visual production in its most basic form — drawing. We test the hypothesis that repeatedly drawing an object differentiates its representation from other objects, making it more perceptually discriminable. This hypothesis is motivated by our previous findings that: (1) a deep neural network model of the ventral visual stream trained purely on photographs also recognized drawings, reflecting a common feature representation of objects across production and recognition; and (2) training people to draw improved the model's ability to recognize their drawings, resulting from reduced feature overlap in the representations of different objects. The current study directly evaluates how learning to draw affects human object recognition, using a categorical perception task sensitive to such representational changes. Participants alternated between drawing two Trained objects (e.g., bed and bench), and did not draw an additional two Control objects (e.g., table and chair). Before and after training, participants categorized morphs of the two Trained and the two Control objects. Insofar as drawing differentiated Trained object representations, resulting in less feature overlap between them, intermediate morphs should be recognized according to the distinguishing features of the dominant object in the morph, and perception should become more categorical. Indeed, learning to draw the Trained objects increased the slope of the psychometric curve fitted to categorization responses, suggesting enhanced perceptual discriminability. This was not observed for the Control objects, nor in a separate experiment in which participants watched others draw rather than drawing themselves. Thus, beyond perceptual experience, the act of drawing itself helps refine our object concepts.
Meeting abstract presented at VSS 2017