知识卡片

普通读书笔记卡 · 7011.g · 生成模型:DeepSeek-V4.1-Flash

偏斜的多义与训练-服务偏斜

理解等级:高级

skew 一词被宽泛用于多种数据问题:底层数据的偏置分布漂移、异常值(尤其意外异常值)、数据解释中的语义违背、缺失特征值(特别是某些特征缺而另一些不缺),以及本应同步的两个变量或数据流失去对应;值得注意的是 ML 系统里反而很少取统计学那层「分布偏置漂移」的含义。最可能引发生产问题的是训练-服务偏斜,即模型在训练与服务中表现的任何差异,常见成因是两处特征定义不同、数据本身有变化或缺口、有时甚至是算法与任务之间的反馈回路。这类偏斜是 ML 系统可避免故障的常见原因,所有简单成因都应成为监控目标。但偏斜的检测手段是模型特定的(至少特定于某架构、配置与用途),没有人知道怎么写一个「替所有模型监控各种偏斜」的函数:我们能发现训练与服务有差异,却未必知道差异的成因或它是否有意义,也无法在不知道数据域与预期覆盖的情况下发现覆盖差异。因此监控系统须通用而灵活,而逐个模型/模型族的监控必须由生产工程师与 ML 工程师紧密协作实现。

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

    • Monitoring and Observability for Models · 查看原书依据

      The skew most likely to cause production problems is training-serving skew, which describes any difference between the performance of your model in training and in serving. Common causes are changes in feature definition between training and serving, changes or gaps in the data itself, or sometimes even some kind of a feedback loop between the algorithm and the task.

    参考来源

    - 位置:《Reliable Machine Learning》第09章《Monitoring and Observability for Models》(源文件:extracted/Reliable Machine Learning/_epub-src/OEBPS/ch09.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:The skew most likely to cause production problems is training-serving skew, which describes any difference between the performance of your model in training and in serving. Common causes are changes in feature definition between training and serving, changes or gaps in the data itself, or sometimes even some kind of a feedback loop between the algorithm and the task.