LaTeX实战

前两天学习了LaTeX的基本操作，准备的开始体验使用LaTeX进行排版的酸爽。我上LaTeX模板网址 http://www.latextemplates.com/选择了一篇article的模板，选择了一片英文文章准备小试牛刀。

%建议黑窗键入texdoc lshort-zh 打开文档后进行查阅
\documentclass[twoside,twocolumn]{article}%columns: 决定了每一页的内容是分成一列还是两列。twocolumn: 两列。sides: 决定了打印的时候是双面打印还是单面打印。twoside: 双面打印

\usepackage{caption}
\captionsetup{font={scriptsize}}%设置图表标题字体大小

\usepackage{indentfirst} 
\setlength{\parindent}{2em}%首行缩进两字符

\usepackage{blindtext} %示例宏包

\usepackage{graphicx}
\graphicspath{}

\usepackage[sc]{mathpazo} % 字体宏集
\usepackage[T1]{fontenc} % 切换字体编码要用到fontenc宏包，具体是啥不清楚
\linespread{1.05} % 行间距
\usepackage{microtype} % 调整字体间距宏包

\usepackage[english]{babel} % 英语特殊排版规则

\usepackage[hmarginratio=1:1,top=32mm,columnsep=20pt]{geometry} % 页边距

\usepackage{lettrine} %生成段落首字母大写的效果

\usepackage{abstract} %摘要
\renewcommand{\abstractnamefont}{\normalfont\bfseries} % 将字体设置为粗体
\renewcommand{\abstracttextfont}{\normalfont\small\itshape} % 将摘要本身设置为小斜体文本

\usepackage{titlesec} %章节格式宏包 
\renewcommand\thesection{\Roman{section}} % 章节用罗马数字
\renewcommand\thesubsection{\roman{subsection}} % 用罗马数字表示子节
\titleformat{\section}[block]{\large\raggedright}{\thesection.}{1em}{} % 更改节标题的外观
\titleformat{\subsection}[block]{\large}{\thesubsection.}{1em}{} % 更改子节标题的外观

\usepackage{fancyhdr} % 页眉页脚格式宏包
\fancyhf{}%清除之前对页眉页脚的设置，重要
\pagestyle{fancy} %所有页都有页眉页脚
\fancyhead[c]{Transfer Learning $\bullet$ June 2020} %定义页眉
\fancyfoot[c]{\thepage} % 自定义页脚文本

\usepackage{titling} % 自定义标题部分

%-------------------------------------------------------------------------------
%	标题部分
%----------------------------------------------------------------------------------------

\setlength{\droptitle}{-4\baselineskip} %长度变量赋值

\pretitle{\begin{center}\Huge\bfseries} % Article title formatting
\posttitle{\end{center}} % Article title closing formatting
\title{Transfer Learning} % 题目
\author{%
\textnormal{Bw Hou} \\[2ex]% Your name
\normalsize{SEU} \\ %机构
}
\date{} % Leave empty to omit a date
\renewcommand{\maketitlehookd}{%
\begin{abstract}
Transfer learning is an emerging field of machine learning.It mainly solves the problem of poor model effect caused by missing training data or different distributions of training data and input data.This paper introduces the classifications of transfer learning from the perspective of what to transfer and introduces the corresponding algorithms on how to migrate.Finally, this paper summarizes the existing problems of transfer learning and the further development of transfer learning. 
\end{abstract}
}

%----------------------------------------------------------------------------------------

\begin{document}

\maketitle

%---------------------------------------------------------------------------
%	正文区
%---------------------------------------------------------------------------

\section{Introduction}

\lettrine[nindent=0em,lines=3]{A}lthough the traditional machine learning in classification,regression and other fields has achieved great success, but there are still many restrictions in the actual scenarios. The reason is that machine learning follows two important assumptions: 1) there are enough training data to get a good model 2) the training data and testing data are in the same feature space and follow the same distribution.

But in reality, not all scenarios follow the above conditions. Many examples have proved this view, such as 1) the WiFi localization problem. Wifi localization data in a specific period and another period usually do not follow the same distribution. The trained model of a certain period will perform poorly when it is used in the another period. (2) the Web-document classification problem. The feature space and data distribution of different Web-document are generally different, which leads that the old Web model cannot perform well in the new site classification. And in the new Web, the lack of training data make the model perform poorly.
For human, we can transfer knowledge from a domain to related domains,such as riding a bicycle and motorcycle. If the machine have this ability, it can transfer the data, model and so on from mature (machine learning works well) domains to immature or unknown domains, which will make the machine learning effect in target domains enhance significantly, so the applications of machine learning will be greatly expanded. According to the above thoughts, the transfer learning can be defined as follows: as a new machine learning method,it can use the knowledge from a domain to solve the problems in the different but related doamins. It relaxes two basic assumptions in traditional machine learning and aims to transfer existing knowledge to solve learning problems in the target domain.

In this survey, I mainly introduces the categorization of the TL in“what to transfer”and introduce the corresponding algorithms in”how to transfer”. The rest of the survey is organized as follows:in the next five chapters, mainly discussing the definition,the categorization of the TL and the corresponding algorithms, inadequacy,future prospect and conclusion.

\begin{figure}[]%h表示here，t表示top，b表示bottom，p表示独立一页 
	\centering%居中
	\includegraphics[scale=0.3]{kt}
	\caption{}%添加名字与编号
\end{figure}

%------------------------------------------------

\section{Definition}

There are two important concepts in transfer learning.

1)Domain:A domain contains two parts,feature space X and distribution$P(x),D={x, P(x)}$.
     
2)Task:A task contains two parts, label space Y and prediction function $F(.),T={Y,F(.)}$ 
                                            
Transfer learning mainly has the following three research questions :1) what to transfer 2) how to transfer 3) when to transfer.

“What to transfer”is to distinguish between the data in the source domain that is related to the target domain and that is not. “How to transfer” refers to the method by which the data in the source domain can be directly or indirectly applied to the target task. “When to transfer”refers to the situation in which a certain transfer method can be applied.

%------------------------------------------------

\section{Categorization of Transfer Learning}
\subsection{Instance-based Transfer Learning}
The main idea of Instance-based TL is to select the data with high similarity to the target domain from the source domain ,according to a certain principle.Then we can transfer these data to the target domain to help train the target model. In this way,we can solve the problem of lack of training data in target domain.This method reweights the source domain data in the target domain by measuring the similarity between the source domain data and the target domain data. The weight of source domain data with large similarity is high, otherwise the weight will be reduced.\\
Dai et al. proposed the TrAdaboost method, which applied the idea of AdaBoost to transfer learning. The main idea of this method is to improve the data weight that is beneficial to the target classification task and reduce the data weight that is not conducive to the target classification task.Finally the upper bound of generalization error is deduced based on PAC theory. Huang et al. proposed Kernel mean-matching (KMM) method to estimate probability distribution. Wu Lei et al proposed a Transfer Sparse Hierarchical Probabilistic self-organizing Graph (TSHiPSOG). In this method, multi-level and multi-granularity representation vectors are extracted from source domain and target domain respectively. Moreover, maximum information Coefficient (MIC) was used to measure the similarity degree of vector between source domain and target domain.Finally the part of the vector in source domain was reused to help train classifier in target domain. Lu Jing et al. combined bagging integration method and clustering algorithm to realize sample transferring. First, different clustering algorithms are used to cluster the same learning data.Then each clustering result is weighted by voting. The bagging integration technology is used to filter the source domain data.Then leave the source domain data that are clustered with the target domain data according to the voting results.otherwise the data are eliminated.In this way they can select the data with high similarity to the target domain. Finally,these data are used to help train the target model.

\subsection{Feature-based Transfer Learning}
Feature-based transfer learning mainly seeks for typical feature representatives between the source domain and the target domain to further eliminate the differences between the two domains so as to realize the cross-domain knowledge transfer. The transfer method can be further divided into feature selection transfer learning and feature mapping transfer learning. 

(1) Transfer learning based on feature selection focuses on how to find out the common feature representation between source domain and target domain, and then use these features to carry out knowledge transfer.

(2) Transfer learning based on feature mapping focuses on how to map the data of source domain and target domain from the original feature space to the new feature space.

Pan et al. proposed Transfer Component Analysis (TCA), which is a typical method. The core of the method is to minimize the distribution difference between different data fields by taking the Maximum Mean Deviation (MMD) as the measure. Blitzer et al. proposed a Structural CorrespondingLearning (SCL). The algorithm can transform some unique features in one space to axial features in all other spaces, and then use machine learning algorithms to classify and predict based on these features. Long et al. proposed to combine instance and feature transfer learning methods by adding transfer Joint Matching (TJM) method to minimize distribution distance. Jing Zhang et al. proposed to train different transformation matrices for source domain and target domain respectively, so as to achieve the goal of transfer learning. Gong et al. proposed GFK in 2012. Through a feature mapping, the source domain and target domain can be transformed to a public space with the minimum distance.
\subsection{Model-based Transfer learning}
Model-based transfer learning refers to the method of finding the parameter information Shared between the source domain and the target domain to achieve the transfer. This transfer method requires the assumption that the data in the source domain and the data in the target domain can share some model parameters.

The TransEMDT method was proposed by Zhao et al. This method firstly constructs a behavior recognition model based on decision tree with labeled data .Then for unlabeled data, use the K-Means clustering method to find the optimization of calibration parameters. Deng et al. are also using transfinite machine learning to do the similar work. Pan et al. used HMM to conduct indoor positioning research under different distribution, aiming at the characteristics of dynamic changes of Wifi indoor positioning under different devices, different time and different space. Long et al. improved the deep network structure. By adding probability distribution adaptation layer, they further improved generalization ability of the deep transfer learning network for big data. At present, most model-based transfer learning methods are combined with deep neural networks.These methods modify some existing neural network structures by adding domain adaptation layer into the network, and then conduct joint training.

\subsection{Relationship-based Transfer learning}
Relation-based transfer learning refers to the fact that there is a certain correlation between source domain data and target domain data. Knowledge transfer can be achieved by establishing a relationship model between source domain data and target domain data.

There are very few researches on the transfer learning method based on relationship.These articles all use Markov Logic Net to explore the relationship similarity between different fields.

%------------------------------------------------
\section{Inadequacy}
(1) The process of transfer learning implies a condition that the target domain and the source domain are related. So how to measure the similarity between the two domains? This problem has not been generally answered yet.

(2) There is a very important problem in transfer learning , that is how to avoid negative transfer. It is caused by the fact that the source domain has little correlation with the target domain .And it will make the target model worse. If the former problem is not solved,this problem will not be solved fundamentally.

(3) Can we further improve the adaptability of machine learning by conducting transfer learning in two less relevant areas?

%----------------------------------------------------------------------------------------
\section{Future Prospect}
(1) To solve the second problem above, multi-source and multi-angle transfer learning can be adopted.The method divides the feature space of source domain and target domain into multiple dimensions and domains. After all,one domain knowledge is limited. So we can try to use knowledge from multiple related fields to help train the model of target domain. In this way, the opportunities to find and learn more favorable knowledge for the target domain can be increased, so as to make transfer learning more stable.It also can effectively avoid the occurrence of negative transfer. For example, emotion recognition based on physiological signals can use EEG signals and various peripheral physiological signals such as ECG,HRV,SC,PPG and EMG to train the model.

(2)To solve the third problem above, the distant domain transfer learning can be used. Through the continuous transfer of several related domains, two domains with little correlation can realize distant domain transfer learning. For example, Professor Yang Qiang's team published a paper” Distant Domain Transfer Learning” at the 2017 AAAI a top-level conference about AI. In that paper airplane can be recognized using a face recognition model.

%----------------------------------------------------------------------------------------

\section{Conclusion}
First of all, it must be recognized that transfer learning is an idea that is proposed to solve a class of problems.Its background is to solve the problem of poor model performance caused by the lack of data or different distribution of training data and input data. This means that it can expand the amount of data in the target domain and combine with the traditional machine learning to solve the above problems. The core of TL is what to transfer,how to transfer, when to transfer.This survey introduces the categorization of the TL on “what to transfer” and introduces the corresponding algorithms on”how to transfer”. Finally, this paper summarizes the existing problems and the further development. In the future, multi-angle learning, multi-source domain learning and distant domain transfer learning, the construction of more effective source domain and target domain correlation model and the reduction of negative transfer may become hot research directions.

%----------------------------------------------------------------------------------------
\end{document}

结论：
（1）使用LaTeX进行排版还是很容易的，只需要把别人的内容进行替换即可。
（2）但是让人头大的是可能会有很多排版出来不舒服的地方，需要静静的修改各种“bug”。
（3）刚开始很多代码看不懂，需要借助LaTeX提供的文档才能明白是什么意思。这样可以大大提高学习效率。