Tips on publishing in NIPS, ICML or any top tier conferences for ML
As you must have already known, this year NIPS has seen an exponential growth in registrations. In the face of overwhelming competition, i think it would be great to log the tips for publishing in a top tier conferences. It would be great if someone can comment on the special care for the papers inorder to get published. This is a very subjective question in nature, so any comments are welcome.
[–]BeatLeJuce 34 指標 1 月前*
It generally pays off to have a sensible story line. Like: don't just say, "we did stuff X and it works well and here is why", but have a compelling argument for why your stuff is awesome. Sell your idea, don't just present it. This is NIPS we're talking about, not a boring science conference.
It always helps to have a theorem or two in there. It doesn't even have to be a good/fundamental one. Some reviewers just really like it when you have theorems.
Math is good. You have to strike a balance between having too much math ("you should submit this to COLT instead") and not having enough. If all else fails, write down the definition of a logistic sigmoid or a softmax. Ideally in an align-block.
Make sure you actually hit the paper limit. ICML accepts 8 pages + 1 of references. If your idea fits on 7.5 pages, it clearly isn't fleshed out enough yet.
Make sure you cite everyone who could be a potential reviewer, no matter how relevant their contributions.
It helps to make your paper look like it comes from some bigshot labs. There are certain papers that some labs always cite / arguments they often bring / words they prefer. It's much harder to shut down a presumably bigshot professor than a nobody.
If you want to write an application paper, make sure it's about audio, images or video. No-one really cares about other application areas, unless your plan was to submit to workshops in the first place.
Do repeat experiments and have confidence values on your results. For a field that does a lot of Bayesian stuff, ML reviewers are very fond of those. If you can't do repeats, just be creative with how you define confidence.
Make sure your work is novel. Meaning, work the word "novel" into the abstract and conclusion. Use the introduction to actually explain why your work is novel (spoiler alert: the papers who actually did the same thing you did have probably done so only as a side-note. Make sure you point out that their paper was about something else entirely).
[–]Evizero 12 指標 1 月前
It is so sad how right you are. This reads like a horror story
[–]posteuphoria 6 指標 1 月前
Two things I would add:
9.1: Your work shouldn't be too novel. Especially not novel for novel's sake. In the same vein as 5, if your paper fits badly into existing literature, which most probably has been written by the same people reviewing your paper, it may provoke a negative bias.
10: Most reviewers will form an initial opinion about your paper chiefly by reading the abstract, the introduction (assuming it fits on the first page), whatever your first equation says, and your figures. Even if later they point out flaws in the paper, a good first impression can easily make the difference between an overall weak accept and strong accept vote.
There's also a whole art on writing the reviewer responses, see e.g. this blogposting.
[–]hughperkins 3 指標 1 月前
Haha this is awesome! :-) plausibly useful, and sufficiently cynical to make me smile :-D
[–]mr_robot_elliot[S] 1 指標 1 月前
Yeah that's a great answer, Would it be better to log into the wiki of this subreddit so that anyone can come back to read such stuff?
[–]ddofer 1 指標 1 月前
You can get non audio/video/text/audio work done, depends on the field. (We got a paper on shallow ML methods for proteins into a NIPS Satellite on ML and computational biology - ASAP).
[–]BeatLeJuce 2 指標 1 月前
Hence the line "unless your plan was to submit to workshops in the first place". Getting a paper into the main conference about non-media applications is very hard (though not impossible; Baldi had one about physics at NIPS 2014) -- Getting into the associated workshops is much easier (especially if there is a specialized workshop like the MLCB).
[–]olaf_nij 7 指標 1 月前*
I think the long-term future is quite likely to be something that most researchers currently regard as utterly ridiculous and would certainly reject as a NIPS paper.
-Geoff Hinton
[+][刪除] (2 下層留言)
[–]dwf 6 指標 1 月前
Is your goal to do good science or to get papers accepted? These two goals are not necessarily aligned.
[–]mr_robot_elliot[S] 2 指標 1 月前
Well, the goal is to get good science accepted :)
[–]vodkagoodmeatrotten 6 指標 1 月前*
Try workshops first. Find ways to co-author. Read https://mluebbecke.wordpress.com/2014/11/21/how-to-write-a-paper/ then http://www.crc.gov.my/wp-content/uploads/documents/elements_of_style_for_journal_articles.pdf thenhttp://hunch.net/?p=22
Yeah thanks for the links. Will look into it.
[–]sleepicat 1 指標 1 月前
Agreed. I'm detecting quite a herd mentality already. Not good for those who want to break out of the pack.
[–]kjearns 3 指標 1 月前
The best strategy is to write a clear and compelling document about something that's very interesting.
[–]mr_robot_elliot[S] -2 指標 1 月前
Could you please comment on the structure of the project, how much time would it generally take for a project to get into matured stage enough to be published. Also, are there any specific resources which can be logged into the wiki of this subreddit
[–]Articulated-rage 2 指標 1 月前
Could you please comment on the structure of the project
Scientific Method
how much time would it generally take for a project to get into matured stage enough to be published
Don't count beans. But also, many months.
Also, are there any specific resources which can be logged into the wiki of this subreddit
why?
[–]kagglerihardlyknower 2 指標 1 月前
dude from your questions here you have no chance, just do something else.
what do you mean by something else?
[–]dwf 2 指標 1 月前
Balloon sales is always a good bet if you live somewhere warm and sunny.
yah probably and apply ML to predict how high it goes according to viscosity blah blah :P
浙公网安备 33010602011771号
[–]BeatLeJuce 34 指標 1 月前*
[–]Evizero 12 指標 1 月前
[–]posteuphoria 6 指標 1 月前
[–]hughperkins 3 指標 1 月前
[–]mr_robot_elliot[S] 1 指標 1 月前
[–]ddofer 1 指標 1 月前
[–]BeatLeJuce 2 指標 1 月前
[–]olaf_nij 7 指標 1 月前*
[+][刪除] (2 下層留言)
[–]dwf 6 指標 1 月前
[–]mr_robot_elliot[S] 2 指標 1 月前
[–]vodkagoodmeatrotten 6 指標 1 月前*
[–]mr_robot_elliot[S] 1 指標 1 月前
[–]sleepicat 1 指標 1 月前
[–]kjearns 3 指標 1 月前
[–]mr_robot_elliot[S] -2 指標 1 月前
[–]Articulated-rage 2 指標 1 月前
[–]kagglerihardlyknower 2 指標 1 月前
[–]mr_robot_elliot[S] 1 指標 1 月前
[–]dwf 2 指標 1 月前
[–]mr_robot_elliot[S] 1 指標 1 月前