Proceedings of the Annual Meeting of the Cognitive Science Society

Learning new categories often alters similarity judgments: items from different categories are judged as less similar, and items within the same category as more similar. While category training may drive such changes, categories often have distributional structure without supervision. This raises the question of whether unsupervised exposure alone can reshape similarity judgments, and by what mechanism. We hypothesized that similarity judgments reflect a learned internal geometry of stimulus space. During unsupervised exposure, learners may infer the manifold structure underlying the distribution of examples, such that similarity depends on path-based relations on the manifold rather than inter-item distance or exemplar density alone. To test this, participants experienced stimulus distributions that differed only in whether they formed a connected or disconnected manifold. Despite equivalent physical similarity, participants judged items as less similar when separated by a topological gap. These results demonstrate that unsupervised exposure is sufficient to alter similarity judgments and reflect sensitivity to inferred geometric structure of stimulus space.