← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 8860.e · 生成模型:GPT-6

权重衰减与二范数惩罚需条件

理解等级:高级

在基本梯度下降中,平方L2惩罚的梯度会产生参数乘性收缩,因而可联系权重衰减。该等价不能不加条件推广到带自适应预处理的优化器或不同衰减实现。原书术语提供直觉,实际训练仍需说明更新公式与惩罚系数。

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

      this regularizer is called weight decay. When p = 1, the regularizer, applied to regression problems, is called LASSO regression. For any p, the effect is to encourage most parameters to be zero, or near zero. When most parameters are zero, the function takes on a degenerate form, that is, a simpler form. For example, if we consider the quadratic hypothesis space θ1x + θ2x2,

    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:摘录中关于权重衰减与二范数惩罚需条件的定义、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:this regularizer is called weight decay. When p = 1, the regularizer, applied to regression problems, is called LASSO regression. For any p, the effect is to encourage most parameters to be zero, or near zero. When most parameters are zero, the function takes on a degenerate form, that is, a simpler form. For example, if we consider the quadratic hypothesis space θ1x + θ2x2,