How to read papers
Recently I have been reading many papers in AI, ML, and DL. One day I was curious if there are some proper ways to read papers. So, I googled it and found this article by S. Keshav is very helpful. Here is the summary of the article1.
Three pass approach
S. Keshav introduced the so-called three pass approach. Even though I don't write papers, this approach seems to be an instructive guide how I am supposed read papers after trying out the three pass approach on papers.
The first pass
This step is basically a "skimming."
How to
- Quick Pass
- Less than 10 minutes
- Read title, abstract, and introduction
- Read section and subsection headings
- Read conclusion
- Glance references
Check 5 "C"s
- Categories: what kind of paper is this?
- Context: what theories have been used as bases
- Correctiveness: if the assumption seems valid
- Contributions: what does this paper contribute to?
- Clarity: is the paper easy to read besides of theories?
The second pass
It is about getting to the main context. You should be able to grasp what this paper was about quite in depth. Yet, it's still possible not to understand completely.
How to
- Read but proof part
- Write down key points on its margin
- Look carefully figures
- Mark unread references for future reading
- Should take up to one hour
Check
- If you can summarize the papers with some supporting evidences
The thrid pass
It is the final step to fully understand papers. The key is to re-implement the paper. Think as if you were the main author of the paper you are reading and make sure if you can make the same assumptions as authors. Basically, you are a reviewer or an author to seriously attack the paper.
- It may take up to 5 hours
- You should be able to re-create the same paper
- You should be able to tell its strong and weakness
Conclusion
Thanks to this detailed guide, I found myself now I feel more confident after reading papers. The key portion of this approach is really the third pass. I found I learn so much by re-generating papers though it takes one full day for me.
Reference
[1] S. Keshav, "How to read a paper," http://ccr.sigcomm.org/online/files/p83-keshavA.pdf [↑]