ML Research: It’s All About the Process
By far, one of the hardest (and most rewarding) parts of grad school is learning how to engage with other researchers’ ideas. Being able to read papers in AI has helped me put together better presentations to present to professors, generate more ideas, and discover more niche, open problems in the field than many of my colleagues — but surprisingly, it’s a skill which none of my professors ever actually took the time to teach.
Today, I want to pull back on the curtain on how I read papers (and by extension, you could also apply this to taking lecture notes, textbook reading, etc.), so you don’t have to start from scratch when getting into research.
Step 1: Set the Stage
The way you start reading a paper is like how you start a workout: by warming up.
In my case, my personal research interest lie at the intersection of computational imaging and neural rendering. But, at one point in grad school I was required to take a course in “statistical learning theory” — aka, an entirely different domain of research!
It’s not a good idea to jump into a paper before you’re really curious about what it contains. In my case, I frankly found this step challenging. Eventually, one thing I found helped motivate my paper-reading to peruse the websites of the authors my professor would assign, such as Kunal Tawar (a Research Scientist at Apple). That’s because even though I went in with the impression that reading about theoretical concepts is dry, Tawar’s work touches millions of devices through his role in privacy-preserving machine learning. Now that’s what I call having an industry impact!
Key Takeway: in your own research adventures, before starting a paper try to take a few minutes to explore the author’s background and the paper’s context. This will help you frame the problem and its significance.
Step 2: Outline Your Notes
OK! So now we’re just juiced up our curiosity levels. How will we organize our notes?
In my personal experience, I like to follow the “Outline Method”, as presented by Thomas Frank here. In particular
- write down headings using an ordered tree-list
- and, along with that write down questions that personally curious for you to try and answer, before diving into the details in full
Of course, for this step you can choose just about any kind of outline that works best for yourself.
Step 3: Decode the Language
If you’ve ever picked up a paper and had no idea what the vocabulary meant, then you’re not alone! Researchers are notorious for using jargon that perhaps not even ChatGPT will recognize (for another example related to statistical learning theory, try asking it to define “the fat shattering dimension” for you, and see what happens :). Therefore in this step, we want to recognize the ambiguity, and take a moment to acquaint ourselves by:
- study the figures + captions
- look up any words in bold/italics/quotes that stand out
In other words, this step involves breaking down complex terms and techniques into manageable components. Visual aids like diagrams and flowcharts can be incredibly helpful in this regard.
Step 4: Capture the Core Contributions
Alright! Now we’re cruising. By now, you are probably starting to see through the darkness of Plato’s Cave a little, but aren’t yet able to see the bright day outside. We’re now ready to fully take in this paper though. So the next steps are to:
- read through the paragraphs in-depth
- summarize to make sure you grasp the intuition and any key insights
The intention here is identify the paper’s core contribution, which could be a novel algorithm, an innovative framework, or a unique insight. Jot down the main problem being addressed, the proposed solution, and how it improves upon existing approaches. Think of this as the heart of the paper that you’ll want to retain in your notes for the long term.
Pro-Tip: one thing I’ve noticed is the most advanced researchers pay special attention to one thing, that particularly garners deep insights about their field: they pay close attention to when the papers in a given domain, start referring to the same idea using different names over time. Going back to the NeRF example: if you watch this talk by Vladen Koltun, and notice how he stresses calling NeRF an implicit representation; and then you move this textbook on NeRF where it is part of a section called “Light Field Imaging and Display” — that, my friend, is a dead giveaway that NeRFs are becoming a topic larger than any single person can learn about all on their own. And that’s exciting, because the chances are the more chances there’ll be to find open problems for you to work on if your own research!
Step 5: Embrace the Details
Next strp — this is going to be disappointing. It is. But the truth is: there is no next step.
“Hold up, Zain — what do you mean, there’s no next step?”
What I mean is, there’s no next step that everyone should follow. What I’m trying to say is, everyone should try to read papers to at least get the core intuition. But after that step, how far deeper you go depends on your personal interests.
For me, there are some papers that are super relevant to a project/course I’m working on — in that case, this is the point where I revisit the sections that discuss the technical intricacies, assumptions, and experiments. I ask myself how each component contributes to the overall solution. I’ll scour the Web to seek additional resources (like online tutorials, lecture notes, or dope Medium articles on places like Towards Data Science) to clarify concepts that seem challenging. But for other papers, this could just be the place where I stop entirely.
Pro-Tip: as a stretch challenge, try to reproduce the results presented in the paper on your own also, using code provided by the authors (or better yet, contributing your own open-source implementation if there is not one already)!
Step 6: Reflect and Relate
After comprehending the paper’s content, take a step back and reflect on how it relates to your existing knowledge. Can you draw connections to concepts you’ve encountered before? Can you answer any of the questions you wrote down in Step 2?
Even better, do you have new questions you want to explore :) ? This reflective step is crucial for synthesizing the new information and expanding your mental framework.
Step 7: Engage in Collaborative Learning
As fun as it can be to read papers (no sarcasm here at all), it is a lot more enjoyable to learn about papers together!
Engage in collaborative learning by joining discussions with peers and mentors. Online forums on Discord, seminars, and workshops are fantastic platforms to exchange thoughts and interpretations. Constructive debates can reveal multiple perspectives and provide a richer understanding of the paper’s nuances.
Pro-Tip: for people interested in a NeRF paper reading group, I’d suggest looking into the Discord server for nerfstudio :)
Step 8: Foster Connections
Lastly, reach out to the authors or researchers whose work you find intriguing. Craft thoughtful emails expressing your questions and/or appreciation for their contribution and your interest in their research. Establishing connections within the research community can open doors to mentorship opportunities, collaborations, and even friendships that will enrich your academic journey.
Conclusion
That’s it. Give me feedback on what you think of this process, want to see an example (perhaps in a future post?) or anything else. Happy reading! 🚀📚🧠
