Models of object categorization reflect multiple categorization strategies.

Robert J. Peters(1), Fabrizio Gabbiani(2), Christof Koch(1).

(1) Computation and Neural Systems, Caltech. Pasadena, CA 91125; (2)
Baylor College of Medicine, Houston, TX 77030.

We investigated models of visual object categorization in humans,
using simple line-drawn objects parameterized along four dimensions
unknown to the subjects. The categorization task required subjects to
learn to classify training exemplars in a 2-AFC paradigm into two
classes linearly separable in the objects' parameter space. Following
training, subjects categorized additional test exemplars, and the
classification probabilities of these test exemplars were used to fit
several categorization models, including exemplar, boundary,
prototype, and cue-validity models. Previous results (Peters et al., J
Cog Neurosci Suppl 2000) showed that exemplar and boundary models
accounted for human categorization performance significantly better
than did prototype and cue-validity models, but were not
distinguishable from each other. Therefore, we considered new factors
that may influence the relative performances of the exemplar and
boundary models. First, we tested configurations of training exemplars
that differed only in the orientation in parameter space of the
boundary separating the two categories. This led to no systematic
changes in the relative performances of the exemplar and boundary
models. Second, we tested configurations of training exemplars that
differed in which features were most important in separating the
categories. It appears that when the definition of categories depends
critically on pairs of features ("meta-features") that are in close
physical proximity and map well onto the x-y image plane, the boundary
model performs as well as the exemplar model. Conversely, when the
categories depend on pairs of features that are physically separated
and do not map directly to the image plane, the boundary model
performs significantly worse than the exemplar model, while the
cue-validity model performs as well as the exemplar model. This
suggests that observers use different strategies depending on whether
they are able to treat the relevant features as a unit.