This year, I volunteered to be a NIPS reviewer, since I want to get the experience of reviewing academic papers. It was pretty easy to volunteer: for every paper submitted to NIPS, they asked that at least one of the authors offer to be a reviewer, and preferably not someone who’s already on their list of known reviewers. I’m clearly not among that elite crowd, and I wanted to review anyway, so it was an easy choice for me.
This year, NIPS had around 2500 (!) paper submissions, according to the number I saw in the Conference Management Toolkit website, which is what NIPS uses to manage paper submissions and reviews. I think the exact number I saw was roughly 2450-ish, but it’s worth noting that papers with authors who match my conflict of interest domains (especially “berkeley.edu”) wouldn’t appear on this list. When I was submitting my paper bids for reviewing purposes, I couldn’t find the paper I submitted (fortunately!).
For obvious reasons, I can’t divulge too much about the papers I’m reviewing. But I can probably safely say this: I got assigned four papers to review in the span of roughly a month (from June 17 to July 17). So I’m reviewing right now, and decided to write this blog post during one of my breaks. Now that I’ve read all my assigned papers and have more-or-less come up with a very high-level draft conclusion about them, here are four thoughts:
Unfortunately, I’m probably deeply qualified to review … none of the four papers. For two of the papers, I understand much of the main ideas, but I still lack intricate knowledge of the math and algorithms they use, and about the most prevalent related work. For two other papers, I’m close to being lost on them. Therefore, I have made this a rule: for each paper, I will allow myself to read one extra reference for background information. One. So effectively, my reading load for this cycle is eight papers, four to be read in-depth and four to be read semi-in-depth. This raises my curiosity: how do people actually carefully review papers? The vast majority of NIPS papers are extraordinarily specialized and technical, and many still have proofs in appendices. I don’t see how it is humanly possible for me to verify the proofs and make sure the math lines up correctly, especially with an internship that already takes up about 50 hours a week. It can take hours just to verify one simple line of equations in a paper, which might require looking through a previous paper reference, checking Wikipedia for additional background information, and then marking down an annoying typo that got me confused. Can an experienced NIPS reviewer please tell me how he or she reviews papers?
Fortunately, the previous thought might not matter because, despite not having gone through all the technical material so far (I’ve only been reviewing a week!) I feel like I can more or less tell how I feel about the papers. For instance, the amount of formalism and “cleanliness” of the paper gives a good clue about its quality. If the paper is rife with typos and sloppy LaTeX, or if the figures look spectacularly low-quality, that’s just asking for a rejection. (Come on, with matplotlib embedded in Jupyter notebooks, people should be able to produce acceptable figures.) Note that by this, I’m excluding language issues that might arise due to non-native English speakers writing the paper. It’s not fair to them.
It’s nice reviewing papers without knowing the identity of the authors. NIPS is double blind, which I believe to be a good thing, and I say this as someone who would probably benefit from an open reviewing process. I don’t want to think about the authors’ identities. I have not actively searched so I’m still in the dark about this for all four of my papers. As a reviewer, I like my identity being protected, so I’m reconsidering my earlier thought about making some of the reviews public (sadly, if this were politics, I’d be accused of “waffling”). Unfortunately, for one of my papers, I have a good feeling about who at least one of the authors is based on previous research, despite how I haven’t been in this field very long (and some would argue, have never been in this field). In addition, two of my other papers have a previous reference which is obviously similar to the current work, and since I ended up using those papers as my “background reference” material and seeing practically word-for-word similarities … author similarities follow. I’ll keep my guesses in mind and then see how they turn out later in August.
Despite my internship, I would bet I have more time on my hands than most faculty for reviewing, because I have fewer competing priorities and because I would be assigned fewer papers (hopefully … did I?). Thus, I will try to write some really detailed suggestions or at the very least, thought-provoking comments/questions, and I think the authors would appreciate that even if their papers don’t get accepted. At all costs, I am not going to be writing those dreaded three line reviews that contribute nothing. Sadly, I frequently see these online at the NIPS proceedings, which makes reviews public for accepted papers. As a result, this period from mid-June to mid-July is going to be super busy for me, since I am doing all my reviewing in nights and weekends. I am really trying to make sure my reviews are helpful.
That’s all for now — I need to get back to reviewing.