This post was published in 2021-09-06. Obviously, expired content is less useful to users if it has already pasted its expiration date.
主要内容:狄利克雷分布/多项式分布/正态分布/各种分布;KL散度
符号∝:https://math.stackexchange.com/questions/652207/what-does-the-sign-propto-mean
𝐴∝𝐵 means that 𝐴 is directly proportional to 𝐵. This means that 𝐴=𝑘𝐵 for some constant 𝑘.
狄利克雷分布的均值和期望https://en.wikipedia.org/wiki/Dirichlet_distribution
多项式分布的均值和期望;https://en.wikipedia.org/wiki/Multinomial_distribution#Expected_value_and_variance
二项式分布的均值和期望;https://en.wikipedia.org/wiki/Bernoulli_distribution#cite_note-:0-2
有一个看起来容易理解的例子:https://www.youtube.com/watch?v=y8XELu_ABl0
结合:http://www.mas.ncl.ac.uk/~nmf16/teaching/mas3301/week6.pdfs
以这个视频为例,首先第一步需要:https://en.wikipedia.org/wiki/Expected_value 其中的 Absolutely continuous case
Normal distribution µ ∼ N (µ, σ2)
定义:https://en.wikipedia.org/wiki/Normal_distribution
根据:https://people.eecs.berkeley.edu/~jordan/courses/260-spring10/lectures/lecture5.pdf
首先Fixed variance (σ2), random mean (µ)是一个最简单的高斯;
根据https://zhuanlan.zhihu.com/p/164377502
这一步看起来需补充:
更详细的计算likelihood的视频:https://www.youtube.com/watch?v=Dn6b9fCIUpM
exponential distribution biased:
推导过程:https://www.statlect.com/fundamentals-of-statistics/exponential-distribution-maximum-likelihood
biased证明1:https://stats.stackexchange.com/questions/489912/showing-bias-of-mle-for-exponential-distribution-is-frac-lambdan-1
biased 证明2:https://math.stackexchange.com/questions/815193/bias-of-maximum-likelihood-estimator-of-an-exponential-distribution,注意这里有【不使用琴生不等式】的强硬证明
exponential distribution在上一个问题里被标记为biased,所以这次要换一种处理方式:
用Gamma distribution作为prior!需要证明为什么用了Gamma distribution以后就变成了conjugate prior,同时求posterior distribution.
发现了一张有趣的图:怎么来理解伽玛(gamma)分布? - 知之的回答 - 知乎 https://www.zhihu.com/question/34866983/answer/60191363
遗留问题:如何在不求posterior distribution的情况下证明一个prior是conjugate prior?
首先根据https://en.wikipedia.org/wiki/Gamma_distribution,exponential distribution是Gamma distribution的一个特殊类型。那么就可以先证明conjugate了!只需要阐述/转换,把Gamma distribution(一般)变为exponential distribution(特殊)就可以证明conjugate了!然后再计算posterior distribution. (好像又不对了)
https://people.eecs.berkeley.edu/~jordan/courses/260-spring10/other-readings/chapter9.pdf
http://halweb.uc3m.es/esp/Personal/personas/mwiper/docencia/English/PhD_Bayesian_Statistics/ch3_2009.pdf
https://courses.cs.washington.edu/courses/cse312/20su/files/student_drive/7.5.pdf
wordpress: 我给mathjax添加了一个隐藏滑动条代码
wordpress: 我对autosave机制进行了一点改动,现在数据库飞速增长!
youtube给我推的广告:https://calcworkshop.com/pricing/
KL散度:
首先是正态分布的均值公式:https://math.stackexchange.com/questions/99025/what-is-the-expectation-of-x2-where-x-is-distributed-normally
有一个视频比较贴近,但是讲的太烂:https://www.youtube.com/watch?v=TNJwYuKjqVM,而且没有用到E(X) .
这个网页:https://stats.stackexchange.com/questions/66271/kullback-leibler-divergence-of-two-normal-distributions,提问者有提到过使用E(X)来解决问题,在这个网页里引出了另一个问答:https://stats.stackexchange.com/questions/7440/kl-divergence-between-two-univariate-gaussians
在中文搜索引擎里寻找【高斯分布kl散度】能获得更多答案!
https://zhuanlan.zhihu.com/p/55778595
https://blog.csdn.net/HEGSNS/article/details/104857277
但是遇到了一个问题:log(x)看起来有点格格不入,它默认是ln(x)吗?为什么很多答案(包括stackexchange)都不使用ln(x)?
google关键词【kl divergence log or ln】
使用autosave_table以后,.sql到.sql.gz的压缩程度越来越高了,所以需要换个命令导入了:
$ zcat myfile.sql.gz | mysql -u root -ppassword mydb
这样可以避免解压出.sql
同时还可以使用 gzip -l 对压缩文件进行压缩率检验:
$ gzip -l mysql_dump.tar.gz
参考:https://stackoverflow.com/questions/2756585/how-do-i-load-a-sql-gz-file-to-my-database-importing/2756591, https://askubuntu.com/questions/907608/how-do-i-get-a-file-sizeoriginal-file-size-within-tar-gz-without-uncompress-i
OCR辅助查找工具
解:先求似然函数fn(rl 0) c eupl一: 1 “!2c2 wi一日)].注意处理类似情况的常见手法(即便是=1高维):(1)将与参数无关的部分略去,将等号变成正比号;(2)尽量提取出关于日的信息,不要带求和号。因为 Za 0)” = n(e一En)’ + (ui En)之j(r1e) X 1 erpl— n (8一En)]. 202