知识卡片

普通读书笔记卡 · 9007.b · 生成模型:GPT-6

对抗训练优化内层最坏值

理解等级:高级

鲁棒目标为 minf E(x,y)[max_(θ∈Θ) L(f(T(x;θ)),y)]。内层寻找规定变换集合里高损失的输入,外层调整模型降低该损失。随机增强与它的关键差别是期望和最大值;近似求解内层时,效果取决于攻击强度,不能把一次未找到攻击当成鲁棒证明。

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

      Here we are applying random transformations to the data, and we want our learner to do well in expectation over these random transformations. Instead of doing random transformations, we could instead consider augmenting with worst case transformations, applying those transformations that incur the highest loss possible for a given training example:

    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:摘录中关于对抗训练优化内层最坏值的定义、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:Here we are applying random transformations to the data, and we want our learner to do well in expectation over these random transformations. Instead of doing random transformations, we could instead consider augmenting with worst case transformations, applying those transformations that incur the highest loss possible for a given training example: