吴恩达dalao教你如何读论文（翻译）

前言声明

本文翻译自如下的文章，暂未获得原作者授权翻译，非官方
仅供个人研究学习使用，欢迎转载，请注明本文链接：https://blog.csdn.net/fish_your_god/article/details/107473323

Reproduction
How You Should Read Research Papers According To Andrew Ng (Stanford Deep Learning Lectures)
（译者注：在全中国最大的学习网站bilibili 你能搜到很多吴恩达讲述如何阅读论文的课程录播）

谁是Andrew Ng？(吴恩达) （如果你不知道的话）

• 斯坦福大学教授
• Deeplearning.ai and Coursera的co-founder
• 强就对了

如何去读Researching Papers?

用Pose Estimation这一个细分领域为例

1.搜集相关资源

来源可以是研究论文，中长篇文章，博客，视频，github仓库

1. Assemble collections of resources that focus on the subject matter. Resources can come in the form of research papers, Medium articles, blog posts, videos, GitHub repository etc.

在这个阶段，没有对于资源的限制，只要你觉得重要就都可以留下来
别忘了做一个list

2.深入研究资源

2. In this next step, you will conduct a deep dive of any resource you deem relevant to the subject matter. It is crucial that there’s a method to track the understanding of each shortlisted resources. Andrew Ng, suggests a table of resource plotted against your understanding level that looks similar to the table below.

接下来将深入到相关领域去进行研究。很重要的是，可以用如下方法来跟踪每一个list中的资源
Andrew Ng（大佬）建议采用一张资源图来表示当前的理解程度（如下图）

建议确保您浏览添加到列表中每篇论文的至少10％至20％的内容； 这样可以确保您在已识别的资源中可以接触到足够的介绍性内容，并且能够准确地评估其相关性。最终你会对于有一些文章有深刻的认识并仔细阅读他们。

多少篇论文就够了呢？

According to Andrew, an understanding of 5–20 papers will showcase a basic understanding within the subject matter, perhaps enough understanding to progress to implementation of techniques.
50–100 papers will primarily provide you with a very good understanding of the domain.

• 5-20篇 可以简要的理解这个方向在干什么，能对一些方法进行复现
• 50–100 篇幅 可以让你很深入的了解领域

当你完成了这些阅读，之前的表格会变成这样：

3.自己的注释

3. The third step is a quick tip from what I’ve observed works for me when trying to understand research papers. The third step is to take structured notes that summarises the key discoveries, findings and techniques within a paper, in your own words.

用自己的话来写一些能够概括文章中重要的发现，技巧的结构性注释

阅读一篇论文

逐字逐句从头读到尾并不是理解论文的最佳方式（译者注：我在读的时候就是这么来的TAT）
至少给论文过三遍，才能对于内容有一个不错的把握

第一遍——title，abstract，figures

4. In your first pass, start with reading the following sections within the paper: title, abstract and figures.

第二遍——intro，conclusion，再过一遍figures和文章其余内容

5. The second pass entails you reading the following sections: introduction, conclusion, another pass through figures and scan through the rest of the content.

介绍和结论的部分通常很清晰与简洁，能够有效概括，并且忽略了所有的补充信息。
作为读者这是有益的，因为你获取了这些核心的信息能够继续进行进一步的阅读。

第三遍——全文，跳过复杂的数学、技巧、公式、术语、定义（你不熟的）

6. The third pass of the paper involves reading the whole sections within the paper but skipping any complicated maths or technique formulations that might be alien to you. During this pass, you can also skip any terms and terminologies that you do not understand or aren’t familiar.

可能还要来几遍——关注一些数学，技巧，和不知道的术语

7. Those conducting in-depth research into a domain can take a few more passes. These additional passes will mainly be focused on an understanding of the maths, techniques and unknown terminologies presented within the paper.

Questions To Ask Yourself

在阅读文章的时候，作者提出要问问自己的问题。
这些问题可以告诉关键信息

1. Describe what the authors of the paper aim to accomplish, or perhaps did achieve.
   描述作者期望达成什么，或者已经达成了什么？

2. If a new approach/technique/method was introduced in a paper, what are the key elements of the newly proposed approach?
   如果一个新的方式/技巧/方法在文中提到，那么这个新方法的关键是什么？

3. What content within the paper is useful to you?
   文章中的哪些东西对你有用？

4. What other references do you want to follow?
   你还知道什么其他的参考内容？

一些给出的研究资源

The Machine Learning Subreddit
The Deep Learning Subreddit
PapersWithCode
Top conferences such as NIPS, ICML, ICLR
Research Gate

译者的一些废话：努力就完事了 ^^_ !!!