← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 9004.d · 生成模型:GPT-6

平均风险不约束最坏情形

理解等级:中级

经验风险最小化优化训练样本的平均误差;少量特定扰动上的严重失败可以被平均值掩盖。鲁棒训练尝试优化规定扰动集合内的最坏损失。最坏情形的保障总相对于该集合和攻击假设,不能从一种半径或攻击下有效推广为任意输入都可靠。

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

      Adversarial analysis is worst case analysis. Remember that we typically train our systems to minimize empirical risk, defined as the mean error over the training data. Because we only focus on the mean, the worst case might be arbitrarily bad. Alternative approaches try to minimize worst-case error, and this can result in systems that are more adversarially robust. In chapter 36 we will discuss how

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第35章《数据偏差与分布漂移》(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/c035b.xhtml) - 结论依据:摘录中关于平均风险不约束最坏情形的定义、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:Adversarial analysis is worst case analysis. Remember that we typically train our systems to minimize empirical risk, defined as the mean error over the training data. Because we only focus on the mean, the worst case might be arbitrarily bad. Alternative approaches try to minimize worst-case error, and this can result in systems that are more adversarially robust. In chapter 36 we will discuss how