← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 8983.c · 生成模型:GPT-6

能量似然梯度含正负两相

理解等级:高级

对数似然参数梯度为−E数据[∇θEθ(x)]+E模型[∇θEθ(x)]。做梯度上升时,降低数据处能量,同时提升模型抽样处能量;第二项来自配分函数导数。它避免只降低全部能量的退化,但需要能代表当前模型的样本。

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

      Plugging equation (32.23) back into equation (32.16), we arrive at: (32.24) Both expectations can be approximated via sampling: defining x(i) ∼ pdata and , and taking N such samples, we have (32.25) (32.26) The last line should make clear the intuition: we establish a contrast between data samples and current model samples, then update the model to decrease this contrast, bringing the model closer to

    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:摘录中关于能量似然梯度含正负两相的定义、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:Plugging equation (32.23) back into equation (32.16), we arrive at: (32.24) Both expectations can be approximated via sampling: defining x(i) ∼ pdata and , and taking N such samples, we have (32.25) (32.26) The last line should make clear the intuition: we establish a contrast between data samples and current model samples, then update the model to decrease this contrast, bringing the model closer to