Shepard (1987) observed that generalization gradients exhibited universal properties that were invariant between species and stimulus dimensions, and proposed that these invariances reflect knowledge of "natural kinds" that exist in the world. Here, we test a natural extension of Shepard's ideas, by examining whether generalization differs between different types of learned associations. To do so, we leverage generalization as a tool to diagnose the underlying content of learning in two domains (predictive learning and category learning). In two large experiments equating stimuli and procedural details between tasks, we show that empirical differences in gradients are largely explained by differences in the testing procedures (Experiment 1), and the degree to which outcomes and categories are encoded as mutually exclusive (Experiment 2); not the type of association (predictive vs. categorical) that is learned. Critically, our findings demonstrate that qualitative differences in gradient shape corresponding to rule- and similarity-based generalization are affected by testing procedure, but do not differ by learning domain. We conclude that predictive and category learning are subsumed by the same basic mechanisms, and that the differences in generalization observed between domains reflect knowledge about "cognitive kinds"; categories are created to be mutually exclusive while outcomes are not.
Schlegelmilch et al. (Tue,) studied this question.