知识卡片

普通读书笔记卡 · 7004.c · 生成模型:DeepSeek-V4.1-Flash

管线数据敏感性

理解等级:中级

ML 管线与普通数据处理管线的关键差异:它对输入数据异常敏感,尤其对数据分布的细微变化。只要缺失的那一小部分数据不是随机的、或在模型敏感的特征维度上采样不均,管线就能从“基本正确”滑向“严重错误”。因此聚合、处理与监控数据(而非仅关注在线系统)才是管好 ML 数据管线的关键;监控要点是沿多轴对数据切片(slicing),及时发现某子集整体消失。

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

    • Data Management Principles · 查看原书依据

      All data processing pipelines are, in some sense, subject to the correctness and volume of their input data, but ML pipelines are furthermore sensitive to subtle changes in distribution of the data. A pipeline can easily go from mostly right to significantly wrong simply by omitting a small fraction of the data, provided that small fraction is not random, or is somehow not evenly sampled in the range of characteristics our mod

    参考来源

    - 位置:《Reliable Machine Learning》第02章《Data Management Principles》(源文件:extracted/Reliable Machine Learning/_epub-src/OEBPS/ch02.xhtml) - 结论依据:原文用“丢失某国/某语言数据”说明非随机缺失如何扰乱模型,并指出监控需按轴切片。 - 原始内容:All data processing pipelines are, in some sense, subject to the correctness and volume of their input data, but ML pipelines are furthermore sensitive to subtle changes in distribution of the data. A pipeline can easily go from mostly right to significantly wrong simply by omitting a small fraction of the data, provided that small fraction is not random, or is somehow not evenly sampled in the range of characteristics our mod