About reading research papers吴恩达

how to read one papers?

一种阅读论文不好的方法是从从论文的第一个字开始读,一直读到论文的最后一个字。

吴恩达的真实生活,走到任何地方,都在双肩包里放着一些要阅读的论文。他领导着一个阅读团队,每周主持两篇论文的讨论。为了选择出这两篇论文,可能要阅读6-7篇论文才能选择出其中两篇。

 

阅读方法:

take multiple passes through the paper:

step 1: read the title/bastract and also the figures.

step2:read more carefully the intro, the conclusions, look carefully the figures again and skim the rest.

人们发表学术论文时,发表的流程是 试图是审稿人确信,你的论文值得被接受。另外你会发现,摘要和结论是作者认真总结他们工作的地方。

skim the related work section 跳过相关工作的介绍。有些作者为了使得稿件能通过,大量引用审稿人以前发表过的论文,所以这部分价值部是很大。

step 3: skip the math 跳过数学部分

step4: read the whole thing but skip part that don't make sense.

 

另外,当你读论文并试图理解这些论文的时候,需要问自己下列的问题:

1,what the authors try to accomplish?

2,what were the key elements of the approach?

3,what can you use yourself?

4, what other reference do you want to follow?

 

deep learning一直在发展,如何能获取最新的发展资料:

1,twitter a, kiankatan, Andrew Y Ng,关注牛人

2, ML subreddit

3,顶会:NIPS, ICML, ICLR,当这些会议举办时,去浏览下收集的论文,发现自己感兴趣的论文,并使用。

4,friends,可以形成一个团体来分享论文等。

 

对于里面有很多数学的论文,例如batchnorm,如果想深入理解:

step1:read through it, take detailed notes and see if you can rederive it from scratch(不看做好的笔记)。

通过从头开始推导别人的工作,你可以学会如何推导自己的新的算法。

吴在博士时,做了很多这样的事情,

 

对于代码的建议:

I think the simple lightweight version one of learning would be to download and run the open source code if you can find it, and a deeper way to learn this material is to re-implement it from scratch.

step1:run the source code

step2:re-implement it (重现)(完成重现表明你确实理解了算法)

 

Longer term advice:

for users keep on learning and keep on getting better and better , the more important thing is for you to learn steadily (稳定地,稳固的,有规则的。)

坚持每周读两篇论文,而不是一个假期读50篇。

 

你可能感兴趣的:(深度学习)