← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8982.c · 生成模型:GPT-6

能量需要有限配分函数

理解等级:中级

能量模型定义pθ(x)=exp(−Eθ(x))/Z(θ),其中Z是整个定义域的积分或求和,需满足0<∞。能量越低表示相对密度越高。任意神经网络分数不自动满足可归一性,训练时不显式计算Z也不意味着可忽略分布存在条件。

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

      Density models constrain the learned function to be normalized, that is, ∫x pθ(x)dx = 1. This constraint is often hard to realize. One approach is to learn an unnormalized function Eθ, then convert it to the normalized density , where is the normalizing constant. The Z(θ) can be very expensive to compute and often can only be approximated.

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第32章《生成模型》(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/c032.xhtml) - 结论依据:摘录中关于能量需要有限配分函数的定义、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:Density models constrain the learned function to be normalized, that is, ∫x pθ(x)dx = 1. This constraint is often hard to realize. One approach is to learn an unnormalized function Eθ, then convert it to the normalized density , where is the normalizing constant. The Z(θ) can be very expensive to compute and often can only be approximated.