Created: December 08, 2021
Modified: December 08, 2021
Modified: December 08, 2021
how to keep up with papers
This page is from my personal notes, and has not been specifically reviewed for public consumption. It might be incomplete, wrong, outdated, or stupid. Caveat lector.- People on Reddit worry that there are hundreds of new ML papers every day---how could you possibly keep up? How can you filter the firehose into what's important and relevant?
- My answer: don't try to "keep up" with papers. It's not the right goal. Most papers are insignificant, wrong, redundant, poorly written, or some combination of these. Even among award-winning papers, for any given researcher only a tiny fraction will be relevant, and the majority of these will still be forgotten in a few years. The firehose of papers is not the fountain of wisdom that you think it is. The vast majority are worthless, and even the gems are valuable only to a very specific subset of people which probably doesn't include you. So don't worry about missing out.
- How should you find papers to read?
- Paper-reading is not random access. You need context and structure. You need to be able to fit the paper into your larger conceptual scaffolding, to understand how it relates to previous work and what problems it's trying to solve, and to judge how well it actually solves them. Two good sources of structure are:
- Course websites: often an expert will have carefully curated a reading list of important papers.
- Citation graphs:
- Prioritize fundamentals. Most 'new' papers are overly-complicated extensions of some simple set of foundational ideas. Find a few papers in an area, look at the few citations that come up repeatedly, and then read those papers.
- Even better: read textbooks. These condense the lessons of many papers, are usually better written, come with built-in structure, and often exercises. Don't get stuck trying to learn everything about everything (reading every page of TAoCP is not the best use of your time), but before trying to learn a topic from papers, you should make sure you've at least read the relevant chapters of the canonical textbooks.
- To keep up with new material:
- Talk to people. Ask what work they're really excited about and why. Conferences are full of manufactured work that not even the authors are excited about. "Does there exist at least one real human who understands and recommends this paper?" is a surprisingly strong filter.
- Watch talks. (many are online).
- Understand that academia is made of communities, and good research doesn't spring from a vacuum. Any given area will have a few groups that regularly put out good work, and your literature review should give you a decent sense of who those are.
- Conferences are also incredibly valuable. By setting aside a week or so to catch up on what's happened recently, you stay current on the newest ideas to within a year or so, which is usually good enough for all practical purposes. (if you're working on stuff that loses relevance inside of a year, maybe reconsider what you're trying to accomplish). And they're full of talks, tutorials, and people to talk to.
- Paper-reading is not random access. You need context and structure. You need to be able to fit the paper into your larger conceptual scaffolding, to understand how it relates to previous work and what problems it's trying to solve, and to judge how well it actually solves them. Two good sources of structure are: