Proceedings of the Annual Meeting of the Cognitive Science Society

Object recognition and categorization is a fundamental aspectof cognition in humans and animals. Models have been imple-mented around the idea that categories are sets of frequentlyco-occurring features. Out of these models a question has beenraised, namely what is the mechanism by which we learn a hi-erarchically organized set of categories, including types andsubtypes? In this paper we introduce such a model, the Domi-nant Property Assembly Network (DPAN). DPAN uses an un-supervised neural network to model an agent which developsa hierarchy of object categories based on highly correlated ob-ject features. Initially, the network generates representations ofhigh-level object types by identifying commonly co-occurringsets of features. Over time, the network will start to use aninhibition of return (IOR) operation to examine the featuresof a categorized object that make it unusual as an instance ofits identified category. The result is a network which, earlyin training, represents classes of objects using coarse-grainedcategories and recognizes objects as members of these generalclasses, but eventually is able to recognize subtle differencesbetween subtypes of objects within the broad classes, and rep-resent objects using these more fine-grained categories.