Comment by mjn | Hacker News Reader

Comment by mjn | Hacker News Reader

Parent Full thread

> Also, why is NEURIPS or ICML papers is not a hiring guarantee? I thought they're highly sought after.

They're sought after, but the conferences have also grown huge. NeurIPS 2018 accepted around 1,000 papers! Based on a query of the DBLP [1] dataset, there were 4,409 distinct authors who had a paper at either NeurIPS 2018 or ICML 2018 (or both). If you add in a few of the other big AI and ML conferences (AAAI, IJCAI, ICLR), the number grows to 10,995 distinct authors, again solely for the year 2018. The field is hot, but is it hot enough for ten thousand people to be automatically hired because of one paper?

There's also decreasing confidence in the big conferences' review processes I think. NeurIPS 2014 actually did a study to estimate how random acceptance was by assigning some papers to two different sets of reviewers and checking how similar the decisions were [2], and found there was a much higher degree of luck in acceptance/rejection decisions than they had expected. I personally have more confidence in the review processes of smaller and more focused conferences (and journals!), though they don't have the same level of name recognition.

[1] https://dblp.uni-trier.de/

[2] http://blog.mrtz.org/2014/12/15/the-nips-experiment.html

A Ph.D.'s worth of first authored NeurIPS/ICML papers will get you a very good job pretty easily still. But AI slices papers very thin and author lists are inflated relative to other subfields of CS. A single paper in one of the major conferences is a pretty marginal contribution, especially if you're in the middle of a long author list.

Also, NeurIPS reviewing has gone to absolute hell. I mean, peer review everywhere has problems. But I've never seen something quite this bad. At this point I think it's safe to say that most reviewers wouldn't even make it to an on-site interview for a faculty position at a research university. That's definitely nowhere near normal. You can't really blame anyone, I guess; the community is growing way too quickly for any real quality control.

Frankly, I think those conferences have outlived their usefulness as anything except marquee marketing events. I'm now mostly attending smaller and more specialized conferences.

I've gone in a similar direction. Only at smaller conferences can you have any kind of confidence that your reviewers are people with actual expertise in the field. That's pretty useful, not only because it makes it less likely you'll get reviews that are very annoying, but also because a review by a knowledgeable person can be genuinely valuable. The big conferences are full of reviews written by 2nd-year grad students, because with this many submissions, any warm body with anything approaching credentials is needed.

Besides just "quality" in the general sense, one thing this has really hurt, I think, is any sense of history or continuity. There are a ton of reviewers who have basically no familiarity with the pre-2010 ML literature, and it kind of shows in both the reviews and the papers that get published. I mean I get that deep learning beats a lot of older methods on major benchmarks, but it's still not the case that literally every problem, controversy, and technique was first studied post-2010.

Reply on news.ycombinator.com