Building BRIDGES : combining analogy and category learning to learn relation-based categories
The field of category learning is replete with theories that detail how similarity and comparison based processes are used to learn categories, but these theories are limited to cases in which item and category representations consist of feature vectors. This precludes these methods from learning relational categories, where membership is determined by the structured relations binding the features of a stimulus together. Fortuitously, researchers within the analogy literature have developed theories of comparison that account for this structure. This thesis bridges the two approaches, describing a theory of category learning that utilizes the representational frameworks provided by the analogy literature to learn categories that may only be described through the appreciation of the structured relations within their members. This theory is formalized in a model, Building Relations through Instance Driven Gradient Error Shifting (BRIDGES), that shows how relational categories can be learned through attention-driven analogies between concrete exemplars. This approach is demonstrated through several simulations that compare similarity-based learning and alternatives, such as rule-based abstractions and re-representation. We then present a series of experiments that explore the reciprocal impact of relational comparison on category structure and category structure on relational comparison. This work provides a theoretical framework and formal model suggesting that feature-based and relation-based categories are a continuum that are learned through selective attention and similarity-based comparison.