知识卡片

普通读书笔记卡 · 8982.d · 生成模型:GPT-6

能量差提供相对密度

理解等级:中级

对同一模型和定义域,p(x1)/p(x2)=exp(E(x2)−E(x1)),配分函数相消。因此相对排序和若干MCMC更新不需显式Z。不同模型间不能未经各自归一化就用绝对能量比较概率,任意加常数也不改变单模型分布。

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

      Energy-based models (EBMs) address this by simply skipping the step where we normalize the density, and letting the output of the learner just be Eθ. Even though it is not a true probability density, Eθ can still be used for many of the applications we would want a density for. This is because we can compare relative probabilities with Eθ:

    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:摘录中关于能量差提供相对密度的定义、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:Energy-based models (EBMs) address this by simply skipping the step where we normalize the density, and letting the output of the learner just be Eθ. Even though it is not a true probability density, Eθ can still be used for many of the applications we would want a density for. This is because we can compare relative probabilities with Eθ: