Learning to categorize objects in the world is more than just learning the specific facts that characterize individual categories. We can also learn more abstract knowledge about how categories in a domain tend to be organized -- extending even to categories that we've never seen examples of. These abstractions allow us to learn and generalize examples of new categories much more quickly than if we had to start from scratch with each category encountered. We present a model for "learning to learn" to categorize in this way, and demonstrate that it predicts human behavior in a novel experimental task. Both human and model performance suggest that higher-order and lower-order generalizations can be equally as easy to acquire. In addition, although both people and the model show impaired generalization when categories have to be inferred compared to when they don't, human performance is more strongly affected. We discuss the implications of these findings.