Proceedings of the Annual Meeting of the Cognitive Science Society

How humans visually cluster points is relevant for perception, information visualization, and many other domains. However, there are relatively few empirical studies and models of this ability. Here, we propose a new competitive clustering model that uses neurally plausible mechanisms such as hebbian learning and lateral inhibition. We evaluate its fit to the data from a behavioral study of visual clustering, as well as the fits of two categorization models (the Rational model and SUSTAIN) and one statistical learning algorithm (K-Means). We find that people are highly reliable in their clusterings of the same stimulus on two occasions, suggesting they are using a stable strategy. The models were generally successful at predicting human clusterings and were able to replicate the qualitative performance profile of human reliability over different numbers of points and levels of cluster structure. The performance of the competitive clustering model motivates further investigation of its computational properties and empirical validity.