Proceedings of the Annual Meeting of the Cognitive Science Society

An issue in higher-order reasoning is the influence of irrelevant surface (perceptual) features in tasks involving a deep(relational) structure. Many machine learning models use feature vector representations of objects. However, the extent towhich these representations predict or explain human behavior and learning is unclear. A feature vector model facilitatesabstraction and transfer when weights on irrelevant features are minimized and weights on the diagnostic (relational)features are increased. The current study tested whether a feature vector model applies to human behavior in the contextof magnitude relations (line ratio comparison). We systematically varied the degree of surface feature variation whilemaintaining relational structure. We found that, consistent with a feature vector model, participants were more accurateat recognizing common relational structure when surface features differed (t = 4.22, p ¡.001). This approach may bepreferable to a progressive alignment approach to relational magnitude learning.