Human visual object categorization is best described by a model with
few stored exemplars

Despite the success of exemplar-based models of categorization, their
full complexity is not captured by their statistical number of degrees
of freedom: the generalized context model (GCM) stores all training
exemplars in memory, whereas a prototype model (WPSM) stores only one
memory trace per category, yet both have the same statistical number
of degrees of freedom. Since natural visual categories include many
exemplars, the GCM's outperformance of the WPSM might rely on an
implausibly dense neural representation. To make memory use more
explicit, we developed a radial basis function "roaming exemplar
model" (RXM[n]). The RXM[n] follows the GCM/WPSM in classifying
exemplars based on a weighted sum of their attention-weighted
similarity to stored exemplars, but departs from the GCM/WPSM in
allowing the stored exemplars to vary in number per category ([n]) and
to be refined through training. Thus, the RXM[n]'s degrees of freedom
reflect its memory use. We asked human observers (N=9) to classify
sets of Brunswik faces (12 sets, each with 2x10 training exemplars for
two categories, and 60 test exemplars). We fitted the RXM[n] (with
n=1-10 stored exemplars per category), GCM, WPSM, and linear decision
bound model (PBI) with individual subjects' data, and assessed the
fits using the Akaike Information Criterion (AIC) to penalize models
with higher complexity. The best-fitting model was the RXM[1],
followed by the PBI and the GCM. Thus, when models' memory use is made
explicit, categorization behavior is best described with few stored
exemplars, which are refined through training. Unlike the dense
representation predicted by the GCM, this result predicts that a
sparse neural representation different from prototype abstraction
underlies categorization.