how-many-pages-should-paper-span.md

How many pages should a (ML/CV) paper have?

This week, the ever thirsty machine learning (ML) / computer vision (CV) community "find & make trendy" a submission under reviewed for ICLR'22.

The format of the papers spanned an opinionated discussion in social media and at least one Slack channel. About the question:

How many pages?

less is more, but ...

We already know that, right? If not, try to attend a talk from a senior researcher (Professor) presenting a collection of papers in 5 minutes or less.

However, a sharp reader might have anticipated that the scope of a talk is way different from submitting a paper for peer-review purposes. IMHO, during peer-review, someone should present an idea such that others can scrutinize it. Namely, (1) decide if the idea complied with the scientific method, and (2) expand the body of knowledge further. Thus, enough rigor in the presentation is needed to fulfill the two objectives.

so, give me a number!

I'm sorry to disappoint you; I don't have a concrete number. Indeed, I do not care about the number of pages as long as:

1. the paper is clear and well-written.

2. I can assess it. Namely, (i) it's easy to traverse, (ii) find enough relevant experimental or theoretical evidence for all the claims.

3. I can fit it on my schedule.

Note that the first two constraints are primarily under the author's control. Some tips and checklists that tell about the quality of the presentation are:

• Use straightforward English.

• Avoid bullshiting. Namely, when throwing conjectures use the right verbs: "suspect", "hyphotesize", "presume", "believe", etc. Instead of factual verbs: "show", "evidence".

• Organize your paper such that we tell a story instead of describing our likely non-linear and chaotic research journey.

  The organizational part and the fact that digital media (PDF) evolves make any concrete number of pages just a guesstimate. TBH, 4-8 are good enough up to single or double-column format.

• Jia-Bin's tip for writing broken!

Do I genuinely buy that a paper doesn't need eight pages?

It depends on what you want.

For respect, I won't comment about the technical aspects of the paper under review. I found the presentation and general content as a cute piece of experimental research with sarcastic tones (title, figure 1, etc.).

I do know that some companies see publications as the lowest-tier for their business growth. Thus, for people working there, writing less than 8-pages is a blessing. Now, is it also a blessing for the game between authors, reviewers & judges?

A lot of things sound good at the abstract level. But, be careful about what we demand. Those who have read about policy-making understand that it's tricky. As with any experimental research (deep learning?), it needs experimentation. However, a policy does not deter the application of a principle, e.g., "less is more."

On impact, and another concern of people with the paper

• Isn't it disrespectful to present an idea onto a "top-tier venue" without the maximum number of pages?

  First, deciding what's disrespectful is up to the judges. If the paper is clear, why would someone punish authors for being creative in presenting their work?

  Similarly, venues become "top-tier" due to impact, not due to the number of pages. Thus, short papers are OK, especially on an extensively covered topic such as image classification in ImageNet and CIFAR.

• Short presentation might correspond to simple ideas. Thus, those are the impactful ones.

  In other words, "More papers like this are needed. Some ideas are simple enough that we need to shrink the number of pages. The judges are good at detecting impactfulness regardless of the number of pages."

  The quote above is another cup of tea. Rating something as "impactful" is as subjective as "is/not novel." I wrote about the latter before.

  My questions to those claiming to be excellent and gauging impact are:

  • Isn't it impact something that we measure in the future?

  • If so, let's consider this case. If I vote for accepting a paper, as I deem it valuable, shouldn't the community assess my decision in the future?

  

  For those who like gambling or believe themselves good gamblers, shall we run a stress test?

  How? read my entry about novelty. I left an idea there for our peer-review system, fractional authorship/contribution 😜.