Getting the Picture

In the early 20th century, Gestalt psychologists challenged prevailing notions regarding human perception. They showed that there is a difference between seeing the pixels that make up a picture and understanding what a picture represents. We have all had that "aha" moment, for instance, where a scene suddenly becomes clear. The more general point is that people may have all of the information needed to draw a conclusion yet--in contrast to standard economic models--they fail to connect the dots. We build a model that conceptualizes this idea. An agent’s task is to learn about a picture's features. Initially, they know the color of each pixel, but they do not know features of larger regions of the picture. In this sense, the agent starts with complete information but no understanding of what the picture represents (e.g. "it's a smiley face.") They add to knowledge by loading existing knowledge into working memory and deducing new features. Importantly, the agent has limited working memory, which bounds their ability to draw conclusions. We show that the model captures a number of important phenomena, such as multi-stable perception, and provides a useful conceptualization of narratives as "big-picture statements." We discuss several potential applications, including to the politics of persuasion.