Abstract
Categorization is a fundamental and innate ability of our brain, however, its underlying mechanism is not well understood. Previous experimental work on human categorization shows data that is consistent with both discriminative and generative (typically Bayesian) classification. However, the experimental designs used were not able to deliver a test that unambiguously accepts one method while simultaneously rejecting the other. Therefore, we designed a novel experiment in which subjects are trained to distinguish two classes A and B of visual objects, while exemplars of each class are drawn from Gaussian parameter distributions, with equal variance and different means. During the experiment, we test how the subject's representation of the categories changes as a result of being exposed to outliers for only one of the categories, A, far from category B (i.e. increasing category A’s variance). Generative classifiers are by necessity sensitive to novel information becoming available during training, which updates beliefs regarding the generating distribution of each class and assumes class A's variance has increased significantly, thereby reaching across the region occupied by B which predicts an emergence of a new boundary. In contrast, discriminative classifiers are sensitive to novel information only if it affects the immediate discrimination of classes. We observe that, initially, when both categories have equal variance, subjects' decision boundary lies between the two categories consistent with both discriminative and generative algorithms. However, the introduction of the outliers for category A, influences the subject's knowledge of the distribution associated with alternative categories such that objects closer to category B and farthest from category A's outliers will be classified as belonging to category A and thus a new boundary emerges, only predicted by our simulations of generative classifiers. These results give evidence that visual categorization is only consistent with generative and not discriminative classification mechanisms.
Meeting abstract presented at VSS 2015