知识卡片

普通读书笔记卡 · 8991 · 生成模型:GPT-6

先验蒙特卡洛容易浪费样本

理解等级:中级

用先验抽z后平均pθ(x|z)可估计pθ(x),但对某个观测有贡献的潜区域可能很小,许多抽样的条件密度几乎为零。问题是提议分布与后验相关区域不匹配,不只是潜维度单独决定的必然结果。

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 previous trick works decently for modeling low-dimensional distributions. Unfortunately, this approach does not scale well to high-dimensions. The reason is that in order for our Monte Carlo estimate of the integral to be accurate, we may need many samples from pz, and the higher the dimensionality of z, the more samples we will typically need.

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第33章《生成建模与表示学习》(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/c033.xhtml) - 结论依据:摘录中关于先验蒙特卡洛容易浪费样本的定义、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 previous trick works decently for modeling low-dimensional distributions. Unfortunately, this approach does not scale well to high-dimensions. The reason is that in order for our Monte Carlo estimate of the integral to be accurate, we may need many samples from pz, and the higher the dimensionality of z, the more samples we will typically need.