知识卡片

普通读书笔记卡 · 8860.a · 生成模型:GPT-6

参数规模不唯一决定泛化

理解等级:中级

参数数量是容量的粗略线索,还需考虑数值精度、参数约束、架构和优化偏好。较大的网络有时可以比小网络泛化更好,因此经典U形示意不能机械预测所有深度模型。反过来,大模型也不因此天然免于过拟合。

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 number of parameters is just a rough proxy for model capacity (i.e., the expressivity of the hypothesis space). A single parameter that has infinite numerical precision can parameterize an arbitrarily complex function. Such a parameter defines a very expressive hypothesis space and will be capable of overfitting data. Conversely, if we have a million parameters, but they are regularized so that almost all are

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第11章《泛化问题》(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/c011.xhtml) - 结论依据:摘录中关于参数规模不唯一决定泛化的定义、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 number of parameters is just a rough proxy for model capacity (i.e., the expressivity of the hypothesis space). A single parameter that has infinite numerical precision can parameterize an arbitrarily complex function. Such a parameter defines a very expressive hypothesis space and will be capable of overfitting data. Conversely, if we have a million parameters, but they are regularized so that almost all are