← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8990.b · 生成模型:GPT-6

高斯解码器构成连续混合

理解等级:中级

若z来自连续先验,解码器把z映射为高斯分量的均值和协方差,则对z积分得到连续混合密度。共享神经参数可定义无限多分量,不需为每个分量单独存参数。协方差需有效;VAE也可使用非高斯先验或条件分布,并非必须高斯混合。

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 function the VAE uses is gθ. For notational convenience, we decompose this function into and to separately model the means and variances of the Gaussians in the infinite mixture. Next we need a continuous domain to integrate our mixture over, and as a simple choice, we will use the unit Gaussian distribution. Then our infinite mixture can be described as:

    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:摘录中关于高斯解码器构成连续混合的定义、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 function the VAE uses is gθ. For notational convenience, we decompose this function into and to separately model the means and variances of the Gaussians in the infinite mixture. Next we need a continuous domain to integrate our mixture over, and as a simple choice, we will use the unit Gaussian distribution. Then our infinite mixture can be described as: