Prototype and exemplar theories of concepts

The classical theory defined a concept by necessary and sufficient conditions. Wittgenstein had already objected that games share no common feature but a "family resemblance" (Wittgenstein, 1953/2009, §§66–67), and in the 1970s Eleanor Rosch made the point experimental: categories have internal structure. A robin is a more typical bird than a penguin; people rate it a better example and verify faster that it is a bird (Rosch, 1975; Rosch & Mervis, 1975).

Two models followed. According to prototype theory, the mind stores a summary of the category (the central tendency of its features) and classifies by similarity to it. According to exemplar theory (Medin & Schaffer, 1978; Nosofsky, 1986), there is no summary: cases are stored, and a new item is classified by its similarity to all of them, weighting features by the attention they receive. Exemplars better explain sensitivity to atypical cases and to correlations among features; prototypes, economy. The classic objection to both concerns compositionality: the prototypical "pet fish" (a goldfish) is neither a typical fish nor a typical pet (Fodor & Lepore, 1996).

Machine learning reinvented both. k-means sums up each cluster in a centroid, a prototype; kNN classifies by nearest neighbors, pure exemplars, what Witten et al. (2017) call "lazy" learning (2017, p. 85). The prototypical networks of Snell et al. (2017) classify new classes from a few examples by distance to the average of their embeddings. Buckner finds the precedent in Berkeley and Hume, who reasoned with particulars standing in for a whole class (Buckner, 2024, p. 105), and Gärdenfors (2000) proposed thinking of concepts as convex regions of a space.

The key difference is that the psychological models took the features and the metric as given; networks learn them. In a language model, is a concept a prototype (a direction, a centroid), a cloud of memorized exemplars, or something neither theory foresaw?