← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8983 · 生成模型:GPT-6

输入梯度不需要配分函数

理解等级:高级

固定θ时,Z不随输入x变化,所以∇x log pθ(x)=−∇xEθ(x)。它给出密度上升方向,可支持优化或带适当随机机制的采样。若对参数θ求导,Z通常依赖θ,不能沿用同样的省略规则。

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

      Knowing relative probabilities is all that is required for sampling (via MCMC), for outlier detection (the relatively lowest probability datapoint in a set of datapoints is the outlier), and even for optimizing over a space of of data to find the datapoint that is max probability (because arg maxx∈X pθ(x) = arg maxx∈X –Eθ(x)). To solve such a maximization problem, we might want to find

    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:摘录中关于输入梯度不需要配分函数的定义、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:Knowing relative probabilities is all that is required for sampling (via MCMC), for outlier detection (the relatively lowest probability datapoint in a set of datapoints is the outlier), and even for optimizing over a space of of data to find the datapoint that is max probability (because arg maxx∈X pθ(x) = arg maxx∈X –Eθ(x)). To solve such a maximization problem, we might want to find