Poster #33 - Predictive Coding Account for the Development of Representational Drawing

• In Event: 3-206 - Poster Session 13
  In Poster Session: PS 13 Section - Cognitive Processes

Integrative Statement

Figures drawn by children provide insights into their cognition. Not only their motor skills but also their cognitive capabilities (e.g., how they perceive the world and how they represent it in their mind) affect the quality of their drawing. Previous studies found that children start scribbling at one year of age and acquire the skill of representational drawing at around three years of age (Cox, 1992; Saito et al., 2011). A study comparing human children with chimpanzees further revealed that representational drawing can be observed only in humans (Saito et al., 2014). Humans tend to complete missing parts of an incomplete drawing whereas chimpanzees never complete them and instead trace the existing parts. An open question here is what cognitive mechanism underlies human drawing.

We hypothesized that the ability of prediction plays a key role in representational drawing. It has been suggested that the human brain works as a predictive machine and tries to minimize the prediction error between incoming sensory signals and top-down prediction (Friston, 2010; Clark, 2013). In order to complete missing parts of a figure, the human brain would first categorize the figure using only the existing parts and then predict the whole figure based on the estimated category while minimizing the prediction error.

We employed a recurrent neural network shown in Figure 1 (Murata et al., 2013) to computationally verify our hypothesis. The network first learns to draw multiple trajectories using backpropagation through time, which is a common method to minimize the prediction error. Multiple trajectories are differentiated by initial states of the context layer, where the initial states corresponding to the categories of the figures are self-organized through learning. Importantly, the network can infer the corresponding initial state to a given figure using backpropagation after learning. Estimating the initial state enables the network to predict the whole trajectory of the figure and thus fulfill the missing parts.

Figure 2 shows experimental results of the network’s drawing. We first trained the network to draw eight types of trajectories with different shapes (ellipse and 8-shape) and different positions (left, right, top, and bottom) (Figure 2(a)). The network was then given an incomplete ellipse on the right side, where only the first 60% of the trajectory was drawn. We compared results under two conditions (Figure 2(b)): (i) the network did not infer an initial state and instead generated trajectories using the average of eight initial states; (ii) the network estimated the initial state using backpropagation and tried to complete the ellipse. The results demonstrate that the ability to infer the initial state enables the network to complete the missing part whereas the network without such an ability cannot. Drawings produced in (i) are similar to scribblings observed in younger children and chimpanzees whereas drawings in (ii) replicate representational drawing observed in older children. These results support our hypothesis that the ability of prediction drives the development of representational drawing.

Authors

• Yukie Nagai, National Institute of Information and Communications Technology
  Presenting Author

• Timothée Anne, École Normale Suprieure de Rennes
  Non-Presenting Author

• Anja Philippsen, National Institute of Information and Communications Technology
  Non-Presenting Author