My perfectly imperfect note-taking system for ML papers

A note-taking system based on Obsidian and Zotero

My approach to reading ML papers had been a constant frustration. Especially during my Ph.D., I tried different systems, but it mostly was a mix of “just remember the paper contents“, “maybe I’ll find this one relevant paper again after an hour of googling”, and “copy-paste the references from my old paper”.

It doesn’t help that ML research moves at such a high velocity.

It was only when I left academia and became a self-employed writer that I finally found a note-taking approach that worked for me. The secret sauce: A system that encourages imperfection.

I’ve read a couple of approaches to note-taking, but none was a 100% fit. However, it was super helpful to see how other people take notes to figure out my own system. I want to share my approach with you, so you can pick the pieces that resonate with you.

the note-taking software I use

Let’s get right to it. On a high level, my software stack is:

Firefox → Zotero → Obsidian

Firefox is for discovering and reading papers. Zotero is the middleman that organizes my references. Obsidian is where I take notes. All three are glued together through plugins.

Here is an example of what the process looks like:

• I’m on Twitter and discover a relevant paper. If I have time, I might read it immediately, or save it for later.

• I start reading the paper in Firefox. Sometimes I download it.

• I might skip the paper, but if there’s any value in it for me, I go to the next step.

• To get the reference into Zotero, I use the “Save to Zotero” plugin for Firefox. It’s one click and the paper is automatically cataloged. If a PDF is available it’s also saved into Zotero. Works like magic.

• In Zotero, the Better BibTeX plugin is running and automatically adds the new paper to an exported bib file which contains all my references. This happens automatically, so I only open Zotero when I have to correct some information.

• The bib file is automatically read by an Obsidian plugin called “Citations”, a plugin that helps with creating literature notes.

• That means to take notes on the paper, I open Obsidian and press “Ctrl-Shift-O”. This shortcut opens my list of papers. I quick-search and select the current paper which opens a note for the paper with meta-information like title and author already filled out.

• Then I usually split my screen, have Obsidian on the left, and the paper on the right. Then I read or skim the paper and write down anything that I need for a project or that I find interesting. Also meta stuff like how the paper connects to my projects and how I judge the quality of the paper.

• I can reference the literature note in other notes as well.

my rules for writing literature notes

Besides software one needs an approach to writing notes. I’ve developed a few “rules” for myself on how to take notes (not super rigid though). These rules are designed to take perfectionism out of my note-taking process.

My “rules”:

• Only use bullet points: It’s a bit ugly and chaotic, but my literature notes are bullet points only. No paragraphs, no headings, and no abstract/intro whatsoever. On this, I’m very strict. Without this rule, I would mix up note-taking and “writing”. And the process is not about writing, it’s about understanding the paper and extracting value from it. As quickly as possible. I try to write full sentences, sometimes I will add formulas and copy-pasted figures, but otherwise, it’s just stream-of-thought jotted down. If my note isn’t a bit messy afterward, I did something wrong.

  

  Example of a longer note

• Length doesn’t matter: A literature note can be 1 bullet point or 100. It doesn’t matter. A note saying “This paper was irrelevant for the SHAP book because they focus on toy data” might be worth a lot when I revisit the paper years later.

  

• Describe the reading status: I have a field called “reading status” in each literature note where I write down how deeply I’ve read a paper. It’s for record-keeping, but it’s also symbolic: Writing “just skimmed the paper” gives me a sense of closure.

• No copy-pasting: Copy-pasting from the paper doesn’t work for me. I have to write it in my own words. Helps against clutter and improves the chances that I understand the note years later. This is a general rule for note-taking that I follow. It’s the difference between nodding along without actual understanding versus forcing yourself to process it.

• Describe the context: Another piece of information I track is the context of why I read the paper. The context field keeps me focused and when I revisit the literature note I know what lens I used when reading the paper.

My note-taking system isn’t perfect but it works well for me. And maybe you’ll find bits and pieces of this process that are worth trying out yourself!