知识卡片

普通读书笔记卡 · 9007.e · 生成模型:GPT-6

规模不替代覆盖质量

理解等级:中级

增加来源、场景和任务可以扩大模型见过的条件,是通用视觉模型的重要训练方向。数据量本身却不能保证任意环境可靠:重复采集同一偏差只会增加已有区域的密度。衡量规模时同时追踪目标条件覆盖、标签质量与跨域表现。

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

    • · 查看原书依据

      The lesson from this chapter is that if you train on more and more data, you can achieve better coverage over all the test cases you might encounter. This raises a natural question: Rather than inventing new augmentation and robustness methods, why not just collect ever more data? This approach is becoming increasingly dominant, and a major focus of current efforts is on just scaling

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第36章《鲁棒性与通用性训练》(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/c036.xhtml) - 结论依据:摘录中关于规模不替代覆盖质量的定义、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 lesson from this chapter is that if you train on more and more data, you can achieve better coverage over all the test cases you might encounter. This raises a natural question: Rather than inventing new augmentation and robustness methods, why not just collect ever more data? This approach is becoming increasingly dominant, and a major focus of current efforts is on just scaling