Mitigating the Curse of Detail: Scaling Arguments for Feature Learning and Sample Complexity