A neural network model of hierarchical category development

Object recognition and categorization is a fundamental aspect of cognition in humans and animals. Models have been implemented around the idea that categories are sets of frequently co-occurring features. Out of these models a question has been raised, namely what is the mechanism by which we learn a hierarchically organized set of categories, including types and subtypes? In this paper we introduce such a model, the Dominant Property Assembly Network (DPAN). DPAN uses an unsupervised neural network to model an agent which develops a hierarchy of object categories based on highly correlated object features. Initially, the network generates representations of high-level object types by identifying commonly co-occurring sets of features. Over time, the network will start to use an inhibition of return (IOR) operation to examine the features of a categorized object that make it unusual as an instance of its identified category. The result is a network which, early in training, represents classes of objects using coarse-grained categories and recognizes objects as members of these general classes, but eventually is able to recognize subtle differences between subtypes of objects within the broad classes, and represent objects using these more fine-grained categories.