<?xml version="1.0" encoding="utf-8" ?><rss version="2.0"><channel><title><![CDATA[djc_1992的博客]]></title><description><![CDATA[]]></description><link>https://blog.youkuaiyun.com/djc_1992</link><language>zh-cn</language><generator>https://blog.youkuaiyun.com/</generator><copyright><![CDATA[Copyright &copy; djc_1992]]></copyright><item><title><![CDATA[[机器学习] 小傻学HMM：嚼烂HMM]]></title><link>https://blog.youkuaiyun.com/djc_1992/article/details/108399768</link><guid>https://blog.youkuaiyun.com/djc_1992/article/details/108399768</guid><author>djc_1992</author><pubDate>Wed, 14 Oct 2020 11:11:51 +0800</pubDate><description><![CDATA[[机器学习] 小傻学HMM：嚼烂HMM1 基本概念介绍1.1 HMM1.1.1 参数定义2.2 二个假设三种问题2 公式推导2.1 question1：Evaluation计算过程：后向算法计算过程：Learning 问题Decoding 问题
1 基本概念介绍
1.1 HMM
马尔可夫链（Markov link）：一种特殊的随机过程，其随机性只与当前状态有关，与过往已发生的状态和将来可能发生的状态都无关
隐马尔可夫链（hidden Markov method）：用来描述一个变化状态是隐藏的，且是离散的马尔]]></description><category></category></item><item><title><![CDATA[[机器学习] 小傻学EM：嚼烂EM算法]]></title><link>https://blog.youkuaiyun.com/djc_1992/article/details/108411462</link><guid>https://blog.youkuaiyun.com/djc_1992/article/details/108411462</guid><author>djc_1992</author><pubDate>Thu, 17 Sep 2020 12:00:07 +0800</pubDate><description><![CDATA[[机器学习] 小傻学EM：嚼烂EM算法EM 算法Jenson（琴生）不等式：E 步，导出Q函数
EM 算法
Jenson（琴生）不等式：
若f为凸函数，有：
f(tx1+(1−t)x2)≤tf(x1)+(1−t)f(x2)
f(tx_1 + (1 - t) x_2) \leq tf(x_1) + (1 - t) f(x_2)
f(tx1​+(1−t)x2​)≤tf(x1​)+(1−t)f(x2​)
其中t∈[0,1]t\in [0, 1]t∈[0,1],同理，若f为凹函数，只需将上式中的≤变为≥即可。\le]]></description><category></category></item><item><title><![CDATA[[nlp] 小傻学数据预处理]]></title><link>https://blog.youkuaiyun.com/djc_1992/article/details/108246351</link><guid>https://blog.youkuaiyun.com/djc_1992/article/details/108246351</guid><author>djc_1992</author><pubDate>Sun, 06 Sep 2020 15:17:52 +0800</pubDate><description><![CDATA[nlp-数据预处理分词分词工具拼写纠错停用词过滤词的标准化
分词
分词工具
Jieba分词 https://github.com/fxsjy/jieba
SnowNLP https://github.com/isnowfy/snownlp
LTP http://www.ltp-cloud.com/
HanNLP  https://github.com/hankcs/HanLP/
拼写纠错
停用词过滤
词的标准化

...]]></description><category></category></item><item><title><![CDATA[[nlp] LDA主题模型]]></title><link>https://blog.youkuaiyun.com/djc_1992/article/details/108336677</link><guid>https://blog.youkuaiyun.com/djc_1992/article/details/108336677</guid><author>djc_1992</author><pubDate>Tue, 01 Sep 2020 14:10:15 +0800</pubDate><description><![CDATA[[nlp] LDA主题模型1 what is LDA?2 LDA构建流程?2.1 LDA生成过程2.2 LDA整体流程2.2.1 字母定义2.2.2 学习过程References
1 what is LDA?
LDA: Latent Dirichlet Allocatio, 文档主题生成模型，也称为一个三层贝叶斯概率模型，包含词、主题和文档三层结构。
生成模型：我们认为一篇文章的每个词都是通过“以一定概率选择了某个主题，并从这个主题中以一定概率选择某个词语”这样一个过程得到。文档到主题服从多项式分布，主题到]]></description><category></category></item><item><title><![CDATA[[nlp] 小傻学transfomer-XL]]></title><link>https://blog.youkuaiyun.com/djc_1992/article/details/108302730</link><guid>https://blog.youkuaiyun.com/djc_1992/article/details/108302730</guid><author>djc_1992</author><pubDate>Tue, 01 Sep 2020 09:01:29 +0800</pubDate><description><![CDATA[[nlp] 小傻学Transfomer-XLwhat is Transfomer-XL?why Transfomer-XL?the architecture of Transfomer-XL循环机制相对位置编码优点不足codeothersReferences
what is Transfomer-XL?
在传统Transfomer中的segment间引入RNN机制：将上一个segment计算好的hidden state进行存储，在计算下一个segment时将上一个segment的这些信息融入到当前segme]]></description><category></category></item><item><title><![CDATA[[nlp] 小傻学百度ERNIE]]></title><link>https://blog.youkuaiyun.com/djc_1992/article/details/108278700</link><guid>https://blog.youkuaiyun.com/djc_1992/article/details/108278700</guid><author>djc_1992</author><pubDate>Mon, 31 Aug 2020 22:27:15 +0800</pubDate><description><![CDATA[[nlp] 小傻学百度ERNIEwhat is ERNIEwhy ERNIE?ERNIE 1.0简介创新点：其他细节：ERNIE 2.0the architecture of ERNIE2.0连续多任务学习多任务训练任务1: 词法级别预训练任务任务2：语言结构级别预训练任务任务3：语法级别预训练任务结论
what is ERNIE
ERNIE：Enhanced Representation through knowledge integration (Baidu)
why ERNIE?

通过实体和短语ma]]></description><category></category></item><item><title><![CDATA[[nlp] 小傻学XLNet]]></title><link>https://blog.youkuaiyun.com/djc_1992/article/details/108315598</link><guid>https://blog.youkuaiyun.com/djc_1992/article/details/108315598</guid><author>djc_1992</author><pubDate>Mon, 31 Aug 2020 22:15:52 +0800</pubDate><description><![CDATA[[nlp] 小傻学XLNetwhat is XLNet?why XLNet?the architecture of XLNetAR &amp; AEAR（Autoregressive LM，GPT）AE（Autoencoder LM，BERT）comparePermutation Language Model双流自注意力codeothers与bert比较References
what is XLNet?
XLNet：Generalized Autoregressive Pretraining for Lan]]></description><category></category></item><item><title><![CDATA[[nlp] 小傻学bert]]></title><link>https://blog.youkuaiyun.com/djc_1992/article/details/108256129</link><guid>https://blog.youkuaiyun.com/djc_1992/article/details/108256129</guid><author>djc_1992</author><pubDate>Fri, 28 Aug 2020 11:19:23 +0800</pubDate><description><![CDATA[[nlp] 小傻学bertwhat is bert?与其他词向量关系why bert?bert general goalthe architecture of bertbert核心思想input &amp; output句子拆分表示why subword？pre-train(Task 1)：Masked LMpre-train(Task 2)：Next Sentence Prediction (NSP）Fine-tuning参数codeothersbert每一层都学到了什么bert变体bert VS GPT]]></description><category></category></item><item><title><![CDATA[[nlp] 小傻学transformer]]></title><link>https://blog.youkuaiyun.com/djc_1992/article/details/108247922</link><guid>https://blog.youkuaiyun.com/djc_1992/article/details/108247922</guid><author>djc_1992</author><pubDate>Thu, 27 Aug 2020 09:00:05 +0800</pubDate><description><![CDATA[nlp-transformerwhat is transformer?attention后续阅读why transformer?the strength of transformertransformer效率提升the architecture of transformerEncodermulti-head attentionwhy multi-headthe architecture of multi-headScaled Dot-Product Attention自注意力计算步骤多头计算说明Decode]]></description><category></category></item><item><title><![CDATA[[机器学习] 小傻学集成学习]]></title><link>https://blog.youkuaiyun.com/djc_1992/article/details/108244744</link><guid>https://blog.youkuaiyun.com/djc_1992/article/details/108244744</guid><author>djc_1992</author><pubDate>Wed, 26 Aug 2020 19:32:19 +0800</pubDate><description><![CDATA[集成学习-Ensemble什么是集成学习集成模型思想集成学习策略Max VotingAveragingWeighted AveragingstackingbaggingboostingBagging VS Boosting
什么是集成学习
集成学习也叫分类器集成，通过构建并结合多个学习器来完成学习任务。
一般结构：先产生一组“个体学习器”，再用某种策略将他们结合起来。结合策略主要有平均法、投票法和学习法等。
集成模型思想
三个臭皮匠顶个诸葛亮。
弱分类器：分类能力不起那个，但其效果又比随机选但效果稍微好一点]]></description><category></category></item></channel></rss>