Abstract
In this paper, a model of fuzzy category formation was presented. This model achieves fuzzy clustering of exemplars and reduction of feature dimensions simultaneously. Human category formation is a unsupervised learning problem. Forming categories is computing optimal partitions from similarity between exemplars. Category formation is thought to be a clustering humans do. Presented model has two assumptions to characterize the model as human category formation. First, humans reduce feature dimensions and choose informative features to form category efficiently. When we observe exemplars, large amount of features are extracted. We can obtain a good result of category formation, if we used all features. However, because of the limited cognitive resources, it is implausible that humans use large amount of features for category formation. Second, human category structure has fuzziness. In many cases, there are no necessary and sufficient defining features to characterize a category. Features shared by many exemplars form categories. From this family resemblance point of view, human category can't stand up without fuzziness. In this paper, reduced k-means method, which achieves crisp clustering and dimension reduction simultaneously, was extended to represent fuzzy category by using entropy regularization method. Then presented model's validity and limitations as a cognitive model are discussed from results of two simulations and rating data analysis.
