← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 9006.e · 生成模型:GPT-6

自举增强没有新增观测

理解等级:高级

若生成器只用 X 训练,再生成 genaug(X) 给学习器 L,整个流程可以合并成新学习器 L̃=L∘genaug,它仍只接收 X。增强可能通过归纳偏置和优化带来改善,但不能解释为凭空获得独立外部信息。原论证不等于证明增强无效。

这张卡片看不懂?

已有反馈 · 0 条

    还没有人提交反馈。

    在整书笔记中的位置

    原书出处 · 1 处

    • · 查看原书依据

      There are some subtleties when working with learned data augmentation. You might be tempted to train gθ on the same dataset you are trying to augment, that is, train gθ on X, then use gθ to augment X. This can have interesting regularizing effects, but it introduces an uncomfortable circularity, which can be understood as follows. We want to create augmented data for training some

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第36章《鲁棒性与通用性训练》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)/_epub-src/OPS/xhtml/c036.xhtml) - 结论依据:摘录中关于自举增强没有新增观测的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。 - 原始内容:There are some subtleties when working with learned data augmentation. You might be tempted to train gθ on the same dataset you are trying to augment, that is, train gθ on X, then use gθ to augment X. This can have interesting regularizing effects, but it introduces an uncomfortable circularity, which can be understood as follows. We want to create augmented data for training some