In Playpen both object categories and relational categories have a graded similarity structure. As in other neural networks, a category is the instances that make it up. When an instance is presented, it results in changes in the weights between units representing correlations between features. Each of these weights is the combined result of all of the instances the network has seen, and each category is realized as a pattern of weights among a number of units. When a new object is presented to the network, it activates object units. To the extent that these units are strongly associated with others through a pattern of learned weights, the object is a good instance of the category represented by those weights and will tend to activate the category label unit.
In the same way, relation instances are better or worse instances of relational categories to the extent that they activate the relation units which are associated with others through the learned weights that make up the relational category. A network trained on instances of HIGHER in which one object is far from the other will fail to activate the HIGHER unit when presented with two objects which are close together.