A Distributed Representation of Multiple Objects in a Visual Scene

Consider the problem of looking at a scene (such as the one shown in Figure 1), and mentally grouping, or structuring, the parts and pieces into a coherent view of the entire picture. Figure 1 could be \parsed" as three squares, two circles and a triangle. It could also be seen as two white and four black objects, as well as a myriad of other, di erent sets of groupings. These competing ways to view the scene can be a ected by low-level attributes (e.g., objects of the same color, objects that are on the bottom, or objects that are close to each other). However one might also group objects using more conceptual categories (e.g., objects that can roll, or objects that are of the type square). Grouping is therefore sensitive to low-level perceptual pressures as well as those from high-level concepts [Hofstadter, 1984]. At rst glance, our simple ability to group objects on these criteria might not seem to be of interest to the cognitive scientist. However, grouping appears to play a role in the formation of conceptual structures, and these structures are believed to be of central importance for many abstract mental processes, such as analogy-making [Gentner, 1983]. A system that is capable of grouping objects in scenes like the one in Figure 1 must be able to process perceptions, concepts, and multiple items. Symbolic systems have di culty handling Figure 1: Given the task of mentally grouping these objects into sets, what groups might one form? People often form groups based on low-level, perceptual information as well as more abstract categorizations.