What papers are you reading? What's your process for reading papers?

Would love to know what papers people are into at the moment.

Also, have you found techniques to help you digest papers efficiently?

4 Likes

I’m reading the following papers by Kamalika Chaudhuri:

  1. https://arxiv.org/pdf/2102.07048.pdf
  2. https://arxiv.org/pdf/2003.02460.pdf

I will be integrating computing the statistics on these papers into my W&B dashboard. Super excited for the results.

2 Likes

Interesting! How do you pick the papers you want to read Diganta? Do you skim a bunch and commit to a few?

For these 2 papers, I actually watched her talk and it struck a cord since I was ideating with a colleague on an idea which is correlated to her work.
I also read papers based on the trending reports at W&B, especially those under the Reproducibility Challenge.

3 Likes

I just read “Omnimatte: How to Detect Objects and Their Effects” and wrote a blog post about it.

My process for reading papers depends on why I’m reading it.

Looking for an interesting read:

  1. Read Abstract, Introduction & Conclusion
  2. Check online if someone has made a blog, YouTube video etc.

Looking to learn about new field / approach:

  1. Follow Google Scholar citations to like find around 5 largely cited papers
  2. Read Abstract, Introduction & Conclusion, in the most cited ones, read the Methodology & Results sections.
  3. Find the best, more recent Related Works section which paints a good picture of the field
  4. Choose a few to spend more time with :arrow_down:

Looking to implement / deeply understand a paper:

  1. Read Abstract, Introduction & Conclusion
  2. Read Related Works
  3. Read Methodology, Results & any blog posts / explainers online
  4. Read the source code if available / any open source implementation

I love hearing about how people find and read research papers.

4 Likes

I am currently reading papers around semisupervised learning. To be specific, reading the SimCLR series.
I have written a blog recently on how to read papers and I specifically read papers using the same approach. It resonates a lot with what @scott shared depending on the level of interest, going about different passes. Hope others find it useful too.

4 Likes

I select papers to read by browsing the trending section at paperswithcode.

For example, the most recent papers I’ve been reading in depth are:

My reading process can be roughly summed up as:

  • I read the abstract online

  • If my interest is roused, I download the paper, read the introduction, skim over the method, and take a look at the figures and result tables

  • Unless I plan on reproducing the paper, I’ll stop at this point.

  • I typically want to reproduce papers when:

    • They propose ‘simple in practice’ yet highly effective methods
    • The experiments aren’t too computationally expensive and can be done on free cloud compute (aka my saviour, kaggle)
    • There aren’t any torch implementations already
    • I have a lot of free time
  • If I do decide to implement the paper:

    • I look for a blog post or a YouTube video that goes over the core concepts. (Yannic Kilcher is often my go to YT channel)
    • I buckle down and thoroughly read the paper, while making annotations and taking notes
    • Very occasionally, there are parts in papers which are a bit confusing, and I can’t find the answer I want after googling. In this case, I send an email to the paper’s authors.
    • Once I have all the theoretical details down, I start working on a kaggle notebook.
4 Likes