← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 8991.c · 生成模型:GPT-6

后验是理想重要性提议

理解等级:高级

对固定x的非负似然积分,理想提议q(z|x)=pθ(z|x)使重要性权重成为常数pθ(x),获得零方差估计。但该后验的归一常数恰是待求似然,通常无法直接评估或抽样。它说明近似方向,不是可立即使用的免费求解办法。

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

      Using the intuition we developed previously, the distribution q we would really like to sample from is the one whose samples maximize pθ(x | z). It turns out that the optimal distribution is q* = pθ(Z | x).15 This distribution minimizes the expected error between a Monte Carlo estimate of the expectation and its true value (i.e., minimizes the variance of our Monte Carlo estimator).

    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:摘录中关于后验是理想重要性提议的定义、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:Using the intuition we developed previously, the distribution q we would really like to sample from is the one whose samples maximize pθ(x | z). It turns out that the optimal distribution is q* = pθ(Z | x).15 This distribution minimizes the expected error between a Monte Carlo estimate of the expectation and its true value (i.e., minimizes the variance of our Monte Carlo estimator).