知识卡片
下界分开重建与先验约束
ELBO=Eq[log pθ(x|z)]−KL(qψ(z|x)||p(z))。第一项鼓励从观测相关的潜样本解释数据,第二项限制近似后验偏离先验。它是减去KL,不能把两项直接都当作需增加的奖励;改为最小化损失时两项符号相应反转。
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
In this form, the VAE objective is presented as the sum of two terms. The first term measures the data log likelihood when the latent variables are sampled from qψ and the second term measures the gap between qψ and the distribution of latent variables, pz, which we actually should have been sampling from to obtain the correct estimate of pθ(x).
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:摘录中关于下界分开重建与先验约束的定义、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:In this form, the VAE objective is presented as the sum of two terms. The first term measures the data log likelihood when the latent variables are sampled from qψ and the second term measures the gap between qψ and the distribution of latent variables, pz, which we actually should have been sampling from to obtain the correct estimate of pθ(x).