知识卡片

普通读书笔记卡 · 9001 · 生成模型:GPT-6

未来分布未必同过去

理解等级:中级

学习理论中的独立同分布假设,把训练与未来评估连接起来;部署环境改变时,这座桥可能失效。低训练误差和同分布测试误差都不能单独保证未来分布上的表现。应说明目标人群、设备、时间与场景,再判断训练和测试覆盖了什么。

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

      The goal of learning is to acquire rules from past data that work well on future data. In the theory of machine learning, it is almost always assumed that the future is identically distributed as the past. In reality, it is almost always the case that this is not true. Many of the failures of machine learning in industry and in society are due to

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第35章《数据偏差与分布漂移》(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/c035b.xhtml) - 结论依据:摘录中关于未来分布未必同过去的定义、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:The goal of learning is to acquire rules from past data that work well on future data. In the theory of machine learning, it is almost always assumed that the future is identically distributed as the past. In reality, it is almost always the case that this is not true. Many of the failures of machine learning in industry and in society are due to