Abstract
Overall shape similarity is an important determinant of shape categorization. We showed previously with parametrically varied shapes that the responses of neurons in monkey inferotemporal (IT) cortex are related to shape similarity (Op de Beeck et al., 2001, Nature Neuroscience). These results agree with spatial categorization models that represent stimuli in terms of inter-stimulus similarity. Here, we investigated more directly how our previous results are related to the sort of stimulus representation in spatial categorization models. We implemented ALCOVE, the standard neural network version of spatial categorization models. In addition, we implemented a new version of ALCOVE, ITCOVE, in which the stimulus representation was composed of the previously reported responses of IT neurons. The performance of both ALCOVE and ITCOVE matched monkey performance, provided (i) that ALCOVE's stimulus representation took into account the differences between perceptual and parametric shape space; and (ii) that the trial-by-trial variance of neuronal responses was taken into account. Nevertheless, shapes are represented fundamentally differently in ITCOVE compared to ALCOVE. Most importantly, the tuning curves of real neurons show much more diversity in properties such as tuning width and tuning regularity. This tuning diversity increases the odds that some neurons differentiate all category exemplars from all other stimuli (category-selective neurons), even if the category rule is highly nonlinear in stimulus space. Such category-selective neurons are selected when ITCOVE learns a category rule, and this selection changes the representational space that underlies categorization. The diversity in tuning properties combined with this selection process could explain why subjects sometimes learn categories in a more flexible way than predicted by categorization models.