Proceedings of the Annual Meeting of the Cognitive Science Society

Exemplar models have been criticized for lacking mechanisms to explain key conceptual phenomena such as the hierarchical organization of concepts. Here, we offer a potential solution. We show that a broad class of exemplar models can be viewed as a special case of global matching models of memory, and that global matching models are themselves discrete-time approximations of Dense Associative Memories (DAMs), a type of recurrent network. Interpreted this way, exemplar models retrieve hierarchical prototypes by modulating competition during retrieval. We demonstrate this ability using artificial data and pretrained GLoVe and Word2Vec embeddings. Our results suggest that exemplar models remain viable candidates for a broader theory of concepts and provide a natural algorithmic account of attractor-like retrieval in the hippocampus, highlighting their relevance in learning theory and cognitive neuroscience.