知识卡片

普通读书笔记卡 · 6557.a · 生成模型:GPT-6

条件VAE的证据下界分两项

理解等级:高级

给定标签c,ELBO可写为E_q[log pθ(x|z,c)]−KL(qφ(z|x,c)||p(z|c))。第一项鼓励解释图像,第二项约束编码分布与先验。它是evidence lower bound;与真实后验的KL出现在下界间隙恒等式中,不能把先验KL和后验KL混成一个训练项。

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

      Our variational inference procedure optimizes a quantity called ELBO, which means expected lower bound on the log-likelihood of the data. Minimizing negative ELBO loss indirectly minimizes reconstruction error and KL divergence between Q(Z|…) and P(Z|…). Pyro implements ELBO in a utility called Trace_ELBO.

    参考来源

    - 位置:《Causal AI》第05章《连接因果与深度学习》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Causal AI/_epub-src/OEBPS/Text/chapter-5.xhtml) - 结论依据:摘录中关于条件VAE的证据下界分两项的定义、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:Our variational inference procedure optimizes a quantity called ELBO, which means expected lower bound on the log-likelihood of the data. Minimizing negative ELBO loss indirectly minimizes reconstruction error and KL divergence between Q(Z|…) and P(Z|…). Pyro implements ELBO in a utility called Trace_ELBO. - 补充参考:[Pyro官方VAE教程](https://pyro.ai/examples/vae.html),model/guide与Inference部分说明共同训练两个模块、优化evidence lower bound;本卡公式按条件变量版本整理。