Learning in Playpen, as in neural networks in general, starts from specific instances. For objects, the first (weak) correlations that are learned correspond to individual objects or small sets of objects, a particular cup or the three cups most often presented. Only after the presentation of a number of exemplars does the pattern of connectivity come to reflect the more general associations we expect for categories such as CUP. This does not mean that specific associations are lost; as long as the network continues to be presented frequently with specific cups, those cups will continue to exist as ``micro-categories'' within the network.
For relations the same context-specifity applies, but now it is the specific object entering into the relation that are relevant. Given repeated presentation of the same doll in the same basket, the network's early representation of CONTAINMENT will be specific to that doll and that basket. When the basket is presented with different dolls in it, the relational concept becomes more general in one way. When different baskets are presented, it becomes more general in another way. Only after considerable variation among both the container and the contained objects would the network come to represent the set of associations characterizing CONTAINMENT (if ever). But the more specific subtypes of CONTAINMENT would not be lost.
Thus in their development and in their endstate, relations in Playpen are bound to the specific objects that were presented during training.