Exemplar theory
We categorize by comparing a new case to specific remembered instances.
What it means
Exemplar theory holds that categories are represented not by an abstracted summary but by the collection of specific instances a person has encountered and stored, and that a new item is classified by its aggregate similarity to those remembered exemplars. It naturally accounts for sensitivity to within-category variability, correlated features, and exceptions that pure prototype models discard, and it predicts effects of particular old items on new judgments. Formal versions, such as the generalized context model, weight stored exemplars by similarity and have fit categorization data closely. The cost is heavy memory demands and questions about how exemplars are selected and retrieved at the moment of decision. Most contemporary researchers favor hybrid views in which prototype-like and exemplar-like processes both operate depending on task and learning.
How classification works
In the formal machinery, a probe is compared against every stored exemplar at once. Similarity falls off as an exponential function of psychological distance, so a near-identical memory counts heavily while a distant one barely registers. Selective attention stretches or compresses each feature dimension, letting a learner emphasize whatever separates the categories in question. The summed similarity of the probe to each category's exemplars then feeds a ratio rule, yielding a probability of responding rather than an all-or-none verdict. Because the decay is steep and the attention weights are tunable, the same stored set can produce a crisp boundary or a graded one depending on how the space is weighted. This is the engine that lets the generalized context model track trial-by-trial choices.
What the evidence shows
Medin and Schaffer (1978) taught artificial categories defined over four binary dimensions; across geometric forms and schematic faces, their context model beat an independent-cue account, chiefly because participants were sensitive to feature correlations the additive model discards. Nosofsky (1986) tied identification and categorization together, fitting a similarity space from confusion data and then predicting categorization with the same parameters. Further signatures favoring stored instances include the old-item advantage, where previously seen items are classified faster and more accurately than equally typical new ones, and the pull a single unusual training case exerts on later judgments. These are lab findings built on artificial, low-dimensional stimuli, so their reach into rich natural categories is often assumed rather than directly demonstrated.
The prototype debate and hybrids
Exemplar and prototype models frequently mimic one another, so telling them apart demands carefully engineered designs. Smith and Minda (2000) argued that for large, well-structured categories, and especially early in learning, a prototype model fits better, implying that abstraction does real work. Vanpaemel and Storms (2008) went further, treating pure exemplar storage and pure prototype abstraction as endpoints of a continuum and finding that people often sit somewhere in between. Neuroimaging has pushed toward coexistence rather than a winner: Bowman, Iwashita and Zeithamova (2020) recovered prototype-like representations in ventromedial prefrontal cortex and hippocampus and exemplar-like ones in lateral parietal and inferior frontal regions within a single task. The working consensus is now pluralist.
Limits and open questions
Three problems recur. Storage: retaining every instance seems implausible, though sparse, decaying, or clustered memories soften the objection. Retrieval: models rarely specify how the relevant exemplars are located at the moment of decision, effectively treating access as free. Generalization: critics long held that instance memory cannot generalize, but machine-learning results show that similarity-weighted exemplars generalize well, blunting that charge. The deeper worry is identifiability. With enough free parameters, an exemplar model can absorb data a prototype process generated, and the reverse holds too, so a close fit is weak evidence about the underlying representation. Progress increasingly comes from converging behavioural, neural, and developmental measures rather than from goodness-of-fit contests alone.
Examples
A radiologist judges an ambiguous scan as malignant because it resembles several specific past cases that turned out to be cancer, not an abstract average tumor.
A birder calls a scruffy juvenile a herring gull because it matches particular birds seen at the harbour last winter, not because it fits an average gull that exists nowhere.
A teacher marks a borderline essay a B because it reads like two specific essays she marked B last year, exceptions and all, rather than against any abstract standard.
A judge decides a new dispute by matching it to two specific prior cases it closely resembles, reasoning from those remembered rulings rather than from an abstract statement of the doctrine.
A sommelier calls a blind pour a Barolo because it evokes a handful of specific bottles tasted before, quirks and all, not an averaged template of the region that exists nowhere.
First described in Medin & Schaffer (1978); Nosofsky (1986).
Key references
- Bowman, C. R., Iwashita, T., & Zeithamova, D. (2020). Tracking prototype and exemplar representations in the brain across learning. eLife, 9, e59360. doi.org/10.7554/eLife.59360
- Vanpaemel, W., & Storms, G. (2008). In search of abstraction: The varying abstraction model of categorization. Psychonomic Bulletin & Review, 15(4), 732-749. doi.org/10.3758/PBR.15.4.732
- Smith, J. D., & Minda, J. P. (2000). Thirty categorization results in search of a model. Journal of Experimental Psychology: Learning, Memory, and Cognition, 26(1), 3-27. doi.org/10.1037/0278-7393.26.1.3
- Nosofsky, R. M. (1986). Attention, similarity, and the identification-categorization relationship. Journal of Experimental Psychology: General, 115(1), 39-57. doi.org/10.1037/0096-3445.115.1.39
- Medin, D. L., & Schaffer, M. M. (1978). Context theory of classification learning. Psychological Review, 85(3), 207-238. doi.org/10.1037/0033-295X.85.3.207