These five facts are not easily alignable with the kinds of relational representations summarized in Table 1. What kind of system can develop different relational concepts depending on its linguistic experiences in the world? Why does development go from being more object centered to more relational as a function -- not of maturation or age -- but of experience with that specific relation? How are the relevant object properties discovered and represented for different relations and how do these yield a graded category structure?
Past approaches all concentrate on the specification of how the objects in a relation are related to one another, and all accomplish this by some explicit relation term (or explicit role name) together with a mechanism for binding the objects to the roles of the relation. But none of these approaches tells us where the relation term or roles come from.
The five empirical facts presented above suggest a new approach, one which seeks an explanation of the substrate and processes out of which relational categories such as ABOVE are formed. At the center of this approach is the question of how specific relation instances are handled. A relation instance is an explicit association of a particular pair of objects, for example, the spatial relation between a book and a table above which the book is suspended. Just as object categories such as BOOK are generalizations over instances of the category, relational categories such as ABOVE are generalizations over instances of the category (see Kersten & Billman, 1997). In the next section, we present a new model of relational representation based on these ideas: Gasser & Colunga's Playpen model [Gasser ColungaGasser Colunga1997,Gasser ColungaGasser Colunga1998].