ICML 2012 Survey Results
Comments about the conference format
the loss is more. i have seen many borderline papers in proceedings.
If ICML wants to make a real-world impact you need to consider applications papers. See Kiri Wagstaff's paper.
Plenary sessions in the morning, Parallel sessions in the afternoon, Posters in the evening
Increase the number of reviewers and make sure each paper is reviewed by people who are experts in the problem addressed by each paper. I know it is very difficult task, but ICML is considered the top ML conference and should remain as is in the years to come.
Posters are the way to go. Follow the lead of NIPS but then move ahead by making the poster sessions during the day (signalling they are important). Many talks are lousy (even if the paper is good). So I suggest you _substantially_ reduce the number of talks.
Another option is to make all of the presentations 15 minutes instead of 20. There is far too much noise in the reviewing process for it to be useful in determining the relative quality of different papers amoung those that have been accepted. Also, lower scores can sometimes indicate problems with writing and have nothing to do with the quality of the ideas and whether they are worth being presented for longer than 5 minutes.
Have the authors upload a 20 minute video of their presentation/results/slides to an appropriate site. Such sites can be an ICML 2012 youtube/vimeo/whatever channel or something with more structure like videolectures where content is categorized by sessions. This should be optional for the authors.
nips like spotlight + poster sessions might be good. Moreover paper acceptance rate might be decreased slightly to 22~25%
No 'short talks'. Either no talk, spotlight (one or two slides, one or two minutes, top), or full talk. I conjecture that the five-minute 'short talk' format will tempt too many people into cramming too much into their slides and/or going overtime. (Perhaps poll again after the conference.)
It is preferable that the length of the talk for each accepted paper is the same.
With the high number of papers presented in parallel, it might be an interesting option to offer all presenters the option of putting up a poster in a dedicated poster area during a lunch/evening session. That way, delegates could either discuss a talk they heard durring the day in more detail, or catch up on talks they missed in parallel sessions. (informing presenters that these posters might be simply print-outs of the slides, rather than full-fledged posters)
+ videotape all talks
I particularly like the ICML decision in the past to give each paper a talk. This allows nice focused sessions, contrary to conferences such as NIPS, CVPR, ICCV.
no idea what would be a good thing to do.
The fact that ICML offers the opportunity to give a talk (even short) to *all* papers is absolutely essential. As a speaker or attendee in ICML, it makes a big difference in the conference experience.
I think this feature should be saved.
Not my problem, but good luck with it.
Move to fewer tracks (5 is too many) and present some as posters only.
Reduce presentations to 17+3 or 16+4 instead of 20+5.
In case of the poster option, I would suggest to have long poster sessions in order to not penalize the papers accepted only as posters.
Shorten each talk a little. Remove the option of 'invited application papers' again.
Make most papers into posters.
more parallel tracks
Each paper has 10 mins and a poster
definitely 5 tracks is enough, perhaps too many. I'd rather add a day and reduce number of tracks while keeping paper count the same.
Create a numerus clausus: authors cannot co-author more than X submissions. This
will force some authors, who tend to flood co-authorship, to limit their submissions, work better on the papers submitted, and given the 'prestige' of ICML, this should make them select their best works to submit (and not 'test the water' with some of them).
Accept papers on a rolling basis throughout the year, without a specific cap on the number of accepted papers. All papers accepted appear equally in the proceedings/online archive. Allow, but do not require, all accepted papers to present a poster and/or short talk, possibly in one of many parallel sessions (with the understanding that there is no added credit for doing this, once the paper is accepted anyway). Select a much smaller subset of papers for presentation, with more plenary sessions, or perhaps in sessions with less parallel tracks.
Short talk + poster is the best idea. I have such experience at ACL related conferences and it can be even valuable having very short talk. The ICML feature of presenting poster in any case is a big plus!