DL学习笔记16 Structured Probabilistic Models for Deep Learning

最新推荐文章于 2024-02-28 18:41:48 发布

feitianlzk

最新推荐文章于 2024-02-28 18:41:48 发布

阅读量671

点赞数

本文链接：https://blog.csdn.net/feitianlzk/article/details/79184026

版权

本文是关于Structured Probabilistic Models的阅读笔记，探讨了如何使用图形描述概率分布结构，以及结构化建模在深度学习中的优势。笔记涵盖挑战、模型结构、采样方法等内容，如定向模型、无向模型、能量模型以及Gibbs采样等。

摘要由CSDN通过智能技术生成

challege of building large-scale probabilistic models
how to use a graph to describe the structure of a probability distribution
approaches of learning about dependencies(which variables need to be able to interact directly)
specific approches to graphical modeling

goal of deep learning:
- scale machine learning to the kinds of challenges needed to solve AI
- understand high-dimensional data with rich structure
classification
- discards most of information in the input
- ignore many parts(structure) of the input
tasks
- density estimation: derive distribution using given samples
- denoising
- missing value imputation
- sampling
lookup table
- one probablility value per possible outcome
- Memory, statistical efficiency, cost of inference, cost of sampling
- we exlicitly modeling every possible kind of interaction between every possible subset of variables
most variables influence each other only indirectly
- omit third, indirect interaction
- $P(CB | A) = P(C|BA)P(B|A)$
  如果 $P(C|BA) = P(C|B)$
  则有 $P(CB|A) = P(C|B)P(B|A)$
  假定 $C$ , $B$ , $A$ have c,b,a cases
  then $P(C|BA)$ require