Writing Advice for Fledging Machine Learning Researchers

General writing protips

• DO: Write before you even do any experiment.
  • Your initial draft = hypothesis
  • Experiments = observations
  • Update your draft based on your experiments = posterior
  • Rinse lather repeat
• DON’T: Leave everything to the last minute, esp. writing.
• DON’T: Wait until the last minute to submit your paper.
  • Server overload is a thing.
• DO: Read widely and a lot (in English!): novels, nonfictions, popular science, etc.
• DO: Take pride of your work and paper. It’s your life’s work.
  • Haphazard, low-effort, inconsistent-styling paper signals to the reader that you yourself don’t care.
  • Why should they care about your work then?
• DO: See paper-writing as a craft. Always hone your craftmanship.
• DO: Obsess over styling (see below).
• DO: Read math books and appreciate the typography & styling.
• DO: Have a blog site. Publish several in-depth blog posts per year.
• DON’T: Be afraid of people think you’re dumb based on your blog posts.
  • See my blog post archive from 2016 — they’re so simple and embarassing.

LaTeX protips

General

• One line = one sentence
  • This will make debugging easier, due to how LaTeX shows errors
• Quotation marks: Instead of "something", write ``something'' (2 backtics & 2 standard ticks)

Styles

Figures

• Must always fill the full paper width (or column width).
  • Use my library.
  • For figures with multiple subplots, avoid creating a single pdf file for each subplot. Instead, follow this
• If you think they are not appropriate for full width, use wrapfig.
  • See also how to do this with my plotting library.

Tables

• Always use booktabs instead of standard table. See this.
• Like figures, must also always fill the full width. See how.
• Also, like figures, use wraptable if your table is not appropriate for full width.

Maths

• Math macros: Use this and add your own often-used math definitions
  • So, gone are the days where you need to write \mathbf{x} again and again. Instead you could just write \vx.

Bibliography

Citations

• Always use natbib! Two ways of citing:
  • ”… has been done before [4].” — in this case, you use \citep and write ... has been done before \citep{someone2024}.
  • “Someone et al, 2024 has done …” — in this case, you use \citet and write \citet{someone2024} has done ...

References / bibtex

• Rule to live by: Dirty, inconsistent, low-effort reference list signals to the reader that you don’t care. Why should they care about your work?
• Don’t just copy-paste from Google Scholar! Always recheck & edit!
• E.g. make sure the proper capitalization:
  • Instead of: title={Introduction to Bayesian optimization}
  • Write this: title={Introduction to {B}ayesian optimization}
  • I.e., always surround the character that needs to be capitalized with { }
• Venue precedence if a publication appears in multiple venues (top = most prioritized):
  1. Journal
  2. Conference
  3. Workshop/symposium
  4. ArXiv/preprint
• For well-known ML conferences, simply use their abbreviations.
  • E.g. “Advances in Neural Information Processing Systems 35” -> “NeurIPS”.
• For conference, use @inproceedings.
  • The only fields needed are title, author, booktitle, and year.
  • No need for other things like page, editor, etc.
• For journal/ArXiv, use @article.
  • Use journal instead of booktitle.
  • Additionally, volume and number must be included.
• Example:

@inproceedings{kristiadi2024sober,
  title={
    A Sober Look at {LLMs} for Material Discovery:
    {A}re They Actually Good for {B}ayesian Optimization Over Molecules?
  },
  author={
    Kristiadi, Agustinus and Strieth-Kalthoff, Felix and Skreta, Marta
    and Poupart, Pascal and Aspuru-Guzik, Al\'{a}n and Pleiss, Geoff
  },
  booktitle={ICML},
  year={2024}
}