Beyond independent latent classes: Testing the limits of human flexibility

A hallmark of human intelligence is the ability to flexibly adapt to novel situations. This flexibility relies crucially on the appropriate generalization of previously acquired knowledge. One influential theoretical framework argues that humans organize their knowledge in a collection of latent classes. Humans could then assign any novel situation to one of the latent classes (or construct a new one if it is too dissimilar), and thus generalize based on older knowledge. However, this framework is not sufficiently flexible to explain human generalization. In particular, we argue that at least three important features are missing: dependency, compositionality, and tuning. To empirically test this, we developed a novel behavioral task, which required adapting knowledge across tasks in order to generalize appropriately in the test phase. Across three experiments, dependency, compositionality, and tuning requirements were increasingly added to the basic task. Results demonstrate that humans are sensitive to all three types of structure. We discuss how current models must be extended to capture human generalization.