Rather than directly associating full-blown objects, relation units in
Playpen associate the features of two objects. Thus relations in
Playpen can ``look into'' their component objects in a way that is not
possible in approaches where the components of relations are
essentially symbolic, labels representing already categorized
objects.
In the learning of the SUPPORT relation, for example, the
shape and even the color of the supported object is potentially
relevant to the relational associations that develop in Playpen, but
this is independent of how the supported object is categorized by the
system, that is, what word unit it turns on.
However, because relational categories involve correlations of all sorts, the categories of the component objects may also be relevant. If the objects in a spatial relation instance are categorized by the network (that is, they activate category label units), these categories may play a role in the particular relation units that are activated, just as primitive object features are.