← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8990.a · 生成模型:GPT-6

普通重建未规定潜先验

理解等级:中级

普通自动编码器只要求训练编码能被解码,潜空间可能存在复杂边界、空洞及任意尺度。训练后直接从标准高斯抽z,并没有保证落在编码器使用过的区域。因此要把解码器用作可靠随机生成器,还需为潜空间分布建立或学习一致的模型。

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

      But how do we get this z? One goal of generative modeling is to be able to make up random images from scratch. So we need a distribution from which we can sample different z from scratch, that is, we need pz. An autoencoder doesn't directly give us this. You might ask, what if, after training an autoencoder, you just sample a random z, say

    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:摘录中关于普通重建未规定潜先验的定义、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:But how do we get this z? One goal of generative modeling is to be able to make up random images from scratch. So we need a distribution from which we can sample different z from scratch, that is, we need pz. An autoencoder doesn't directly give us this. You might ask, what if, after training an autoencoder, you just sample a random z, say