19

I come from a signal processing background and I have even published in a signal processing conference with a PhD student when I was an intern, therefore, the publication process isn't unknown to me.

My PhD supervisors insist on publishing at top-tier conferences (NeurIPS, ICLR, CVPR, etc); however, upon researching these AI and Machine learning conferences, I was surprised to discover their low acceptance rates, which can be as low as 20%! Is this common in this field?

I am anxious about my ability to graduate if I am unable to publish at these conferences. Can you offer any advice, blogs, or tips for publishing at these conferences?

1
  • 15
    The number of people submitting to these conferences is extremely high, it is not uncommon to get close to 10k submissions each year, so even with the low acceptance rate 2-2.5k new papers are presented in each conference. And the growth is also crazy: twitter.com/ICCVConference/status/1637566384522637312
    – GoodDeeds
    Commented Apr 11, 2023 at 4:23

5 Answers 5

27

Crucial question: do your supervisor's students regularly publish there?

These are the possibilities:

  1. If so, and if most of their students do, they have a pipeline and know how to get top-worthy publications done reliably.

  2. If not, and students still graduate, it's fine to just aim high.

  3. If the answer is no, but they keep the students from graduating unless they have publications in these venues, you should reflect if you want to stay in this group at all, or if, perhaps, you want to give it a try at least for a round.

1
  • 5
    Thank you for the answer. I would say possibility (2), some of their students publish at these venues but not all of them.
    – CaliceMan
    Commented Apr 10, 2023 at 17:36
16

Conferences have limited speaking slots, limited by time and space. If you get a lot of submissions in any hot field, you get a low acceptance rate.

Note that venues for conferences are contracted for long in advance, much longer than the submission/acceptance cycle.

Things may settle down if the number of venues increases or the field cools. It might also be that a lot of the submissions were of low quality or represented parallel work on some problem and reviewers had to choose from the better papers there.

In any hot field (ouch) you have a lot of competition.

2
  • 1
    I don't understand "or represented parallel work on some problem and reviewers had to choose from the better papers there." How do reviewers know if there is other work that is parallely submitted to what they are reviewing?
    – GoodDeeds
    Commented Apr 11, 2023 at 4:19
  • 2
    @GoodDeeds, not the reviewer, the conference committee makes selections just as a journal editor does. They see all submissions.
    – Buffy
    Commented Apr 11, 2023 at 11:11
9

These are indeed competitive conferences though they might not be quite as competitive as the percentages appear, since some fraction of the submissions are crankery/junk.

Some of my colleagues do publish in these areas and my impression is that they typically resubmit work that doesn't make it to the "top tier" in more accepting venues, and this is not a barrier to graduation. This is really no different than journal publishing in other fields: supervisors may want their students to aim for Nature and Science papers, but very few papers actually make it in to those venues.

3
  • 1
    OP asked about "publishing", not merely "submitting" at these conferences. Of course, if the supervisors do not have a record of their students regularly achieving this, this is just a illusory requirement. Also, while these conferences are no longer the venue of exquisite originality of old, as a reviewer for these conferences, I know that the papers submitted are certainly no crankwork and still highly competitive, even if no longer exciting to read. Commented Apr 10, 2023 at 16:29
  • 4
    @CaptainEmacs It's not up to the supervisor or the student where they publish, only where they submit. I would assume most of the crankery is filtered out by the editors/chairs before making it to you as a reviewer.
    – Bryan Krause
    Commented Apr 10, 2023 at 16:37
  • 1
    Yes, I would agree in theory. But some supervisors are not reasonable. As for crankery, I was told that in one of the top conferences (which I won't name here), the desk rejection was abolished. I do not know (and did not check) whether this is true. However, I heard about that after noticing why the review tasks I received in that year when it supposedly happened were less relevant and, in fact, interesting, than in earlier years. Not outright trash, but just more tedious. So, without hard proof, I at least have some circumstantial evidence that the selection process may have changed. Commented Apr 10, 2023 at 16:42
9

This has always been the case for computer science conferences; some has low double digits acceptance rates, e.g., ACM SIGCOMM comes to mind.

Top researchers are measured by the number of such conferences they have on their resume. In contrast, in the EE discipline for example, IEEE Trans. are the norm.

The low acceptance rate is a function of the limited number of slots and the huge community. Arguably, the computer science community, especially AI, is much bigger than signal processing.

2

While the acceptance rate of venues can change between different fields (as discussed in this study), values ranging between 20% and 30% are seem to be common for top venues not only in AI/ML but also in other fields including signal processing.

Taking the signal processing/ML example specifically, we can compare the acceptance rate of some top venues:

  • IEEE transactions on signal processing: 32% in 2010 (source)
  • IEEE transactions on image processing: 24% in 2012 (source)
  • Signal processing: 16% in 2022 (source)

Those numbers are in the same ballpark as conferences such as NeurIPS and ICLR.

The other answers elaborate well on the reasons for acceptance rates being low, but I would recommend you not focus on these numbers but instead on whether your advisor's students publish there regularly, as suggested in Captain Emacs's answer.

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .