知识卡片
已知潜先验支持从头生成
仅有编码与解码函数不足以规定新样本怎样出现,还需知道可抽样的潜变量分布。给定p(z)与pθ(x|z),可先抽z再生成x。潜分布与训练编码分布不匹配时,解码器可能在未训练区域产生任意结果。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
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 critical thing in most generative models, which is not necessarily true for representation learning models, is that we assume we know the distribution pz, and typically it has a simple form such as a unit Gaussian. Knowing this distribution allows us to sample from it and then generate images via g.
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:摘录中关于已知潜先验支持从头生成的定义、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 critical thing in most generative models, which is not necessarily true for representation learning models, is that we assume we know the distribution pz, and typically it has a simple form such as a unit Gaussian. Knowing this distribution allows us to sample from it and then generate images via g.