← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 8993 · 生成模型:GPT-6

单样本目标不是精确似然

理解等级:高级

给定x从q抽一个z,log pθ(x|z)是重建期望的一次随机估计;配合解析KL可估计ELBO。它不是精确log pθ(x),也不是未经权重校正的重要性似然估计。ELBO在期望层面为下界,单次随机实现不必逐次低于真实对数似然。

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 other term, the likelihood term, will be approximated by sampling. For each datapoint x, we will take just a single sample from this expectation, as this is often sufficient in practice. To do so, first we encode x to parameterize qψ(Z | x). Then we sample a z from qψ(Z | x). Finally we decode this z to parameterize pθ(X | z), and we

    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:摘录中关于单样本目标不是精确似然的定义、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 other term, the likelihood term, will be approximated by sampling. For each datapoint x, we will take just a single sample from this expectation, as this is often sufficient in practice. To do so, first we encode x to parameterize qψ(Z | x). Then we sample a z from qψ(Z | x). Finally we decode this z to parameterize pθ(X | z), and we