Occam and Bayes in predicting category intuitiveness

摘要

We consider two models of unsupervised categorization, the simplicity model and the rational model. Their comparison is interesting because the models are based on proximal mathematical principles (minimum description length and Bayesian inference), but their implementation is very different (the simplicity model prefers groupings of similar items, while the rational model groupings which have higher utility). The models’ predictions were assessed with a series of artificial datasets, such that each dataset was designed to reflect a simple intuition about human categorization processes. In the case of linearly separable categories, such that each category was composed of two subgroups, and in the case of non-linearly separable categories, the predictions of the simplicity model and rational diverged. Implications for future developments in unsupervised categorization are discussed.