Toward a technology of generalization: how many exemplars are sufficient?