知识卡片

普通读书笔记卡 · 7008.b · 生成模型:DeepSeek-V4.1-Flash

公平即无偏:垃圾进垃圾出的真正含义

理解等级:中级

主流话语把 fairness 直接绑在 bias 上:公平即偏差的缺席,偏差即不公平;担忧的出发点是训练数据可能取自带偏的系统、或以带偏的方式采集,于是有了垃圾进垃圾出(garbage in, garbage out)的说法。但作者要求读者特别小心:垃圾进里的 garbage 未必指坏数据,数据本身可以干净、准确,真正糟糕的是由不具代表性的训练数据或算法偏见所产生的结果。也就是说,偏见是结果层面的问题,而非只贴在输入上。更进一步,偏见的根本来源可以出现在 ML 流水线的许多环节:坏数据只是其一,坏的建模选择(特征、目标、损失、阈值)同样会引入并放大偏差。为什么关键:若把公平问题窄化成洗数据,就会漏掉建模与部署阶段新生成的偏见。怎么做:把采集、标注、特征工程、训练、评估到反馈回流整条链路都纳入偏见审查。代价:审查面与跨团队协作成本显著上升,却不做就无从定位偏差源头。

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

    • Fairness, Privacy, and Ethical ML Systems · 查看原书依据

      But you should be careful to realize that the garbage in does not necessarily refer to bad data. The garbage is the outcome that results from unrepresentative training data or algorithmic bias. This is a key point: it’s important to realize that the fundamental source of bias can come from many steps in the ML pipeline, including bad data, but also including bad modeling choices.

    参考来源

    - 位置:《Reliable Machine Learning》第06章《Fairness, Privacy, and Ethical ML Systems》(源文件:extracted/Reliable Machine Learning/_epub-src/OEBPS/ch06.xhtml) - 结论依据:原文明确阐述该主题,并给出定义、实践或反例。 - 原始内容:But you should be careful to realize that the garbage in does not necessarily refer to bad data. The garbage is the outcome that results from unrepresentative training data or algorithmic bias. This is a key point: it’s important to realize that the fundamental source of bias can come from many steps in the ML pipeline, including bad data, but also including bad modeling choices.