← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 6564 · 生成模型:GPT-6

部署决策会改写训练环境

理解等级:中级

模型依据预测采取行动后,行动可能改变后续标签或输入分布。广告挽救原本低收入季度,会让“坏信号”与高收入一起出现;若忽略决策过程直接再训练,学到的关系可能混入旧策略效果。记录行动及其选择机制是分析反馈的起点。

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

      machine learning models to drive or automate decisions. Those decisions do not impact the data in domains like meteorology, geology, and astronomy. But in many, if not most, domains where we want to use machine learning, those model-driven decisions are interventions—actions that change the DGP. That can lead to a mismatch between the model’s training and deployment conditions, leading to problems in the model’s reliability.

    参考来源

    - 位置:《Causal AI》第07章《干预与因果效应》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Causal AI/_epub-src/OEBPS/Text/chapter-7.xhtml) - 结论依据:摘录中关于部署决策会改写训练环境的定义、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:machine learning models to drive or automate decisions. Those decisions do not impact the data in domains like meteorology, geology, and astronomy. But in many, if not most, domains where we want to use machine learning, those model-driven decisions are interventions—actions that change the DGP. That can lead to a mismatch between the model’s training and deployment conditions, leading to problems in the model’s reliability.