← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 8993.b · 生成模型:GPT-6

高斯重建须保留尺度项

理解等级:中级

一维高斯条件负对数似然为(x−μx)²/(2vx)+½log(2πvx)。若vx固定,优化均值等价于缩放平方误差;若vx可学,log方差项不能省,否则模型可只放大方差逃避残差惩罚。多维还需说明协方差结构。

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 only differences from an autoencoder are that 1) we sample a stochastic z from the output of the encoder, 2) the reconstruction error is scaled and offset by the predicted variance of the Gaussian likelihood model, and 3) we add to this term the KL loss defined previously.

    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:摘录中关于高斯重建须保留尺度项的定义、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 only differences from an autoencoder are that 1) we sample a stochastic z from the output of the encoder, 2) the reconstruction error is scaled and offset by the predicted variance of the Gaussian likelihood model, and 3) we add to this term the KL loss defined previously.