← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8993.c · 生成模型:GPT-6

重参数化保留可导路径

理解等级:中级

对对角高斯,可先抽与参数无关的ε∼N(0,I),再计算z=μ(x)+√v(x)⊙ε。固定这次ε后,z对均值、方差网络可导,从而估计期望梯度。它不是对随机数发生器求导,乘的是标准差而非方差。

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 step is known as the reparameterization trick, as it reparameterizes a stochastic function (sampling from a Gaussian parameterized by a neural net) to be a deterministic transformation of a fixed noise source (the unit Gaussian). To optimize the parameters for the encoder, we only need to backprogate through μz and σz, which are deterministic functions of x, and therefore we have sidestepped the need

    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:摘录中关于重参数化保留可导路径的定义、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 step is known as the reparameterization trick, as it reparameterizes a stochastic function (sampling from a Gaussian parameterized by a neural net) to be a deterministic transformation of a fixed noise source (the unit Gaussian). To optimize the parameters for the encoder, we only need to backprogate through μz and σz, which are deterministic functions of x, and therefore we have sidestepped the need