← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8991.b · 生成模型:GPT-6

重要性权重补偿抽样偏好

理解等级:高级

从q(z|x)抽样,可用pθ(x|z)p(z)/q(z|x)平均估计pθ(x)。提议分布需覆盖被积函数有贡献的区域,权重方差也应可控。只挑高似然z而不保留补偿因子,会改变积分含义;少量有效样本并非无条件保证。

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 brings us to the second trick of VAEs: when approximating the likelihood integral for pθ(x), try to only sample z values that place high probability on x, that is, those z values for which pθ(x | z) is large. Then, a few samples will suffice to well approximate the entire expectation. This trick is called importance sampling. It is a general trick for approximating

    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:摘录中关于重要性权重补偿抽样偏好的定义、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 brings us to the second trick of VAEs: when approximating the likelihood integral for pθ(x), try to only sample z values that place high probability on x, that is, those z values for which pθ(x | z) is large. Then, a few samples will suffice to well approximate the entire expectation. This trick is called importance sampling. It is a general trick for approximating