How to read research papers?
Data Science is evolving fast enough that even though you have learned a lot of basic foundations, learned a lot of tricks, and probably know better than any practitioners how to get algorithms to work; when you are working on specific applications, you will need to efficiently figure out the academic literature in order to stay on top of ideas as they develop.
This post talks about how to efficiently read research papers and how to acquire knowledge yourself from academic literature by briefly summarizing Andrew Ng’s method.
Compile a list of papers. This could include academic papers, medium posts, blog posts, GitHub posts, or whatever text or learning resources you have.
Keep a track of the amount of understanding you have on each paper on the list. This helps you have a general idea of the progress of your research.
About 5–20 papers are enough to have a basic understanding of a new topic, and enough to implement the ideas from those papers, but it is not enough for research. You will need about 50–100 papers to have a very good understanding of the topic.
How to read one paper?
Reading from the first to the last word in one go is a bad way to efficiently read a paper. Instead, take multiple passes through the paper.
1. Go through the title, abstract, and the figures
You get a very good sense of what the whole paper is about without hardly reading any of the text in the paper itself.
2. Go through the introduction, conclusion, figures (once again), and skim through the rest of the text
The abstract, introduction, and conclusion are often when the authors try to summarize their work really carefully, to make a case to the reviewers as to why their paper should be accepted for publication. So, the abstract, introduction, and conclusion give a very clear summary of what the paper is all about.
After completing this step, you know if you want to continue with the paper or not.
3. Read the paper but skim/skip the math and the code
4. Read the whole thing and skip the part that doesn’t make sense
After completing these steps, you have a pretty good understanding of the paper, and if you still want you can go through it multiple times until the entire paper starts making sense. However, that is completely optional and depends on your requirement for research.
5. Understand the math (optional)
If you want a better understanding of the math section of the paper, you can re-derive it from scratch.
6. Understand the code (optional)
If you want a better understanding of the code section of the paper, you can run an open-source code (the easier and faster way for a basic understanding of how the code works) or re-implement the entire code from scratch (the time-consuming and challenging way that ensures complete understanding of the code)
After reading the paper, try and answer the following questions:
- What did the authors try to accomplish?
- What were the key elements of the approach?
- What can you use yourself?
- What other references do you want to follow?
If you can answer these questions, you have a pretty good understanding of the paper.
Reading and understanding paper gets better with practice.
The key is to go from the very efficient high information content first and go to the harder material later.
Aim for steady and spaced repetition of readings instead of short bursts or cramming.
Although this post talks about Data Science, the same concept can be applied to other research as well.
