Home page

Deconstructing Deep Learning + δeviations

Drop me an email Format : Date | Title
  TL; DR

Total posts : 86

View My GitHub Profile

Index page

How to read technical papers - A crash course

Imagine this. You have a 6 page paper in front of you. Something with a long and scary title. Maybe some crazy equations. (Yes math oops). Woah now you see endless text, oh lord is that a graph which makes no sense? Yes. Why why why!! Mom come save me I'm scared.

Reading a research paper is challenging for everyone, especially if you are a beginner and the field is already hard enough. Sometimes it feels like you are slowly sinking into a pool. But hey! I am here to offer some life support. Read on and the next time you pick up a crazy paper, you will know the tricks that might take you ashore.

A little note here. There is no "one true way" to read a paper. This is my way and it has worked for me so far.

I will be using the paper below as a demonstration. You can grab the annotated paper and follow along. You do not need to understand this paper if its not in your field. It is just an example.

DeVries, T., & Taylor, G. W. (2017). Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552.

Before we jump in, what do you my dear reader get out of this article? There are two sections here. - Section 1 : An understanding of any paper and why it was written - Section 2: Specific tips for the freaky bits - Taking out what truly concerns you from a scary piece of literature - The ability to not run away crying (Mom I'm okay now)

Section 1 : Absorbing the paper

Now that we have that sorted, here are the steps I follow. (Refer to the brackets if you are following along with the paper) (I have added a lot of specific tips after this section so look at that as well.)

  1. Take a long deep breath. You are going on an adventure.
  2. Release your breath and skim through the paper. Do not read. Just look at it. See if there are any figures, diagrams. What do you think this is about? (I see an image on the first page. Hmm. Looks pretty cool. I can already somewhat tell what the paper might be about)
  3. Read the abstract. Slowly. Try to identify what the paper says. In most cases, you will get a very good idea of what will come after this. Now it is time to see if this paper is worth your time. Yes, it is an investment. Do you find it interesting? Is it something that you need to do?
  4. Identify the main subject. Eg. Is it a study? Or a new technique? Or an improvement on an existing one? Or a review? This is your main context. (In this paper it is a computer vision technique for data augmentation)
  5. Skip to the end of the paper. No dont worry, you will come back. Now read the conclusion and the future scope. Why? Well because no paper is perfect. And now you know what the authors have achieved and ... what they have not. This will help you understand the content much better. And sometimes might give you some ideas of your own. (Page 6 - analysis and conclusion)
  6. Now go back and read the motivation of the paper. Try to find out if you relate to any of it. Is there any way you can link it to what you know? This is important as well because if you ever run into a similar issue, hey! now you know how to solve it. (Page 3)
  7. The next step is the cream of the cake. You read the main technique/implementation/study. This is generally the most detailed section and the effort of ages of hard work. Give it some time. Find out what the authors did that seem new to you - a new perspective if you will. Look at the tips and tricks they have used. If this paper has won a competition, you might be able to identify what makes it so special. (Page 4)
  8. Read about the experiments or how the authors verified their claims. (Page 4 - experiments)
  9. Take a break. Try to think about what you have learnt so far. If you have been taking notes, look at them once more.
  10. Go to the introduction and related work sections. You should look out for the following (page 2,3 : I learn about the previous attempts at this by other authors)
    • what did the authors do to fill in the gaps
    • have you read any of these papers? If not, any paper that sticks out to you?
    • look for new ideas
  11. Take a long deep breath. Pat yourself on the back. You did it!!!
  12. If something did not make sense, look at section 2. And re read the paper.

Section 2 : Tackling the freaky bits

You think. All those points are great. But you never said how to tackle the math? Or the spooky techniques that are going above my head. I have been scammed!! Dont worry, heres tips on how to handle them. Read on.

  1. Math. I can see you freaking out. (Sorry math majors).
    • Learn how to say the greek letters. (https://jakubmarian.com/pronunciation-of-the-greek-alphabet-in-english/) Trust me it is way harder to look at it and see symbols and not know how to speak them. It is really easy!
    • Math looks more complicated than it is. In most cases, the authors painstakingly explain every part of those weird equations. Read them.
    • Write them down separately. Yes it seems weird but try it. Just gather them from around the paper. You will see how they magically fit into each other and seem a lot less scarier.
    • Read them multiple times. The first time you are probably too scared anyway. The second, a little more. The third? The fourth? It gets easier.
    • If in doubt, look it up on google

  2. The actual technical bit

    • Compare it with what you already know. Whats different?
    • Have the authors done something new? Or improved an existing thing you already know?
    • Look for references if you find unknown terms. Videos, pictures anything

  3. Literature survey - Yes its probably huge and scary. But you do not actually need to read the full thing and scream. The point of it is to identify the gaps in what existed and how the authors filled it. - If you follow the steps I showed you, you can easily do this! And then the next time you read the survey, wow! it makes sense :)

  4. Graphs/diagrams etc

    • Do they scare you? Well you are not alone. But they are actually not hard to understand
    • Try to put them in words and you will get a lot further

Section 3: Bonus tips

Yes this was meant to be a small article. But I really think this needed to be said and I wish that anyone who comes across this one get more out of it that I initially intended. So here you go my dear readers, a bonus section.

  1. Annotate while you read! I cannot stress this enough. It gives you focus and helps you find whats important
  2. Take notes! They will solidify your understanding a lot.
  3. If it is something that you can try or implement for yourself, try it. It will help you improve your skills drastically
  4. Reread your notes and annotations
  5. Try to explain it to yourself as simply as possible. (Speak to me like I am 14.)
  6. Do not forget that this is a hard task. With time you get better. So you are doing much better than a lot of people just by following this!

I hope this little crash course helped some of you. I wish I had this when I first started reading papers but well, here you go. Still have questions? Ask me in the comments :)

I believe in you! You've got this. Go read that paper that has been scaring you!

If you want to reach out : https://github.com/SubhadityaMukherjee