In Playpen, learning a category means learning the correlations among a set of features. For both objects and relations, this presupposes the preprocessing required for extracting the features, for example, for detecting the edges that are the boundaries of objects. For the learning of an object category all that is then required is that two or more object features be activated simultaneously. The co-occurrence of a particular texture and a particular color may be the beginning of a category such as YOGURT. For the learning of a relational category, there must also be a set of activated units, in this case relation units, but now there is the additional constraint that the phase angles of the corresponding micro-roles of the relation units match one another. That is, the relation units need to have been activated in response to the features of two objects, which the system must have distinguished from one another.
Note that the network can treat a scene consisting of more than one object as a single object, activating object units and learning feature correlations, but learning nothing relational about the scene because no relation units have been activated.
Because they do not presuppose the object segregation that is required for the learning of relational categories, object categories are easier and earlier than relational categories in Playpen.
The fact that each relation unit brings together two features or sets of features also means that, all else being equal, more relation units than object units are required to cover the space of possibilities. Thus, given the same amount of resources in the network for objects and relations, the space will be covered more sparsely by relation units. This should lead to slower learning of relations.