知识卡片

普通读书笔记卡 · 8835.b · 生成模型:GPT-6

公平目标决定约束是否适合

理解等级:中级

一种去偏方法约束预测分布接近训练数据分布,它依赖训练分布本身符合目标的假设。如果原数据已经不合适,忠实保持分布可能延续问题。公平指标、群体划分和理想分布需要明确理由,不能用某个指标改善证明全部伦理目标达成。

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

      Still another approach to dataset fairness, by [527], injects constraints to require that the model predictions follow the distributions observed within the training data, assuming that the training data has the desired distribution. See also [495] for a benchmark and comparison of techniques for dataset bias mitigation.

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第04章《计算机视觉与社会》(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/c004.xhtml) - 结论依据:摘录中关于公平目标决定约束是否适合的定义、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:Still another approach to dataset fairness, by [527], injects constraints to require that the model predictions follow the distributions observed within the training data, assuming that the training data has the desired distribution. See also [495] for a benchmark and comparison of techniques for dataset bias mitigation.