Tips on Learning and Implementing Research Paper

I use a mix of findmypapers.ai, deep research/ perplexity and google scholar to find an initial set of papers. I also like scouting out for review papers as those will give you a general understanding of the field.

From these initial set of papers, I then go on Google Scholar and see which works cited those papers, and likewise see anything interesting from the related works section. This should give you plenty to work from.

Since you're working on an ML system the first step would be to reproduce their results. This means going on Github, downloading their data+code then running it on your system. Once you've reproduced results then do experiments on a smaller scale and record results. This'll help you iterate much more quickly since training a model will be much faster.

This is when you can swap out each submodule with your own piece by piece. For example you might want to swap out their dataset with yours, or you might want to change hyperparams, or use an entirely different model. The idea is to be as methodical and atomic as possible since there are millions of ways ML systems can go wrong. This process helps you create a fast and tight feedback loop.

If the papers you're reading don't have code/data, then tough luck. It'll be quite challenging to reproduce results.