《Advice on how to read papers》---吴恩达CS230随堂笔记

吴恩达教授在他CS230秋季的课上专门用来一节课介绍如何读一篇文献，个人觉得挺好的，就做了随堂笔记。当你阅读这篇博客时，希望能给迷茫的你带来一点点的灯光。

Reading research papers

• Compile a list of papers.
  You can find them from medium(a website) or some blog posts.[e.g. Github、CSDN].

• Skip around the list.
  
  以一种并行的方式阅读论文，也就是同时处理多篇论文。具体地说，试着快速浏览并理解每一篇文章，而不是全部读完。也许你读了每一篇文章的10-20%，但这足以让你对手头的文章有一个高水平的理解。在那之后，你可能会决定删除其中的一些论文，对那些具有开创性的文献，进行精读。
  同时，他也提到了：
  5-20 papers:你可能对这个领域有了基础的知识，或者可以实现一个简单的系统，但并不足以使你走在前沿。
  50-100 papers：你可能对这个领域有了更好的理解。

• Decide which you want to pay attention to.

Reading a paper

It’s a bad way to read from the first word to the last one. Instead, you should take multiple passes below.

• Read the title, the abstract and the figures
  通过阅读文章标题、摘要、关键网络架构图，或许还有实验部分，你将能够对论文的概念有一个大致的了解。在深度学习中，有很多研究论文都是将整篇论文总结成一两个图形，而不需要费力地通读全文。

• Introduction+Conclusion+figures+skim the rest
  介绍、结论和摘要是作者试图仔细总结自己工作的地方，以便向审稿人阐明为什么他们的论文应该被接受发表。
  此外，略过相关的工作部分(如果可能的话)，这部分的目的是突出其他人所做的工作，这些工作在某种程度上与作者的工作有关。因此，阅读它可能是有用的，但如果你不熟悉这个主题，有时很难理解。

• Read but skim the math

• Read the whole paper but skip parts that don’t make sense.
  With the rapid development of the technology, it is usual that the author did not know what should be important in his paper. So he may also write a lot about something that will not make sense in the future.

Questions to ask yourself after reading a paper

• what did the author try to accomplish?
• what were the key elements of the approach?
• what can you use yourself?
• what other references do you want to follow?

Sources of papers

• Twiter
  找寻论文作者的博客或者网站，他关注的人。从中得到一些资源。

• ML subreddit

• NIPS/ICML/ICLR

• Friends

Resources in the paper

• Math

  • Read through it.
  • Rederive from scratch.
    It will help you totally understand and even generate your own novel algorithms in the future.

• Code

  • Find and run open-source code.
  • Reimplement it form scratch.
    It’s a sign of your total understanding of the algorithms in the paper.

Longer-term advices

• Read regularly, often 2 to 3 papers a week.
• Print paper down.