Machine Learning

Reddit posts are problematic.

I've been an active participant in the Machine Learning subreddit for quite some time now, and lately, I've noticed a trend that's been concerning. While the subreddit serves as an incredible hub for knowledge sharing and discussions around ML, there's a growing issue with the quality and reliability of some posts.

Numerous submissions lack proper context, thorough explanations, or credible sources, making it challenging for newcomers and even seasoned practitioners to discern accurate information from misinformation. This trend isn't just about incomplete explanations; it also extends to the validity of claims made in these posts.

It's important to acknowledge that not all content falls into this category—there are incredible insights shared regularly. However, the influx of hasty, ill-explained, or unverified information is diluting the overall value the subreddit offers to the community.

In a field as intricate as machine learning, accuracy and credibility are paramount. Misleading or incomplete information can misguide newcomers and even experts, leading to misconceptions or wasted efforts in pursuit of understanding or implementing certain techniques.

Thus, after observing this trend over some time, I firmly believe that there is indeed a problematic issue with the quality and reliability of several Reddit posts within the Machine Learning subreddit. It's a plea to the community to uphold standards of clarity, depth, and substantiation in discussions and submissions to maintain the subreddit's integrity and credibility.

The system is quite broken, one could say that in its present state, it almost discourages genuine novelty of thought.

But, it's imperfect, first and foremost, because the people involved are imperfect. Reviewing is often a job assigned to the lowest performers in research groups, or traded by the highest performers (constantly on-big tech internships, building startups/open source models on the side) with their colleagues that have a somewhat more laid-back attitude to research excellence. You can submit a bad review and it will not come back to bite you, but in the age of reproducibility, a messed-up experiment or a poorly written/plainly incorrect paper that slips through the review system could be your end.

The idea is that you enter the publishing game at the beginning of your PhD and emerge seeing through and being above the game once you've graduated. After all, you first have to master the rules of the game to be able to propose meaningful changes. It is just that once done, you might have a lot more incentives to switch to industry/consultancy and not care about the paper-citation game ever again.

I think there are a lot of good reasons to be unhappy about the standards of publishing in CS and ML, but I wanted to highlight this:

Who remembers that job posting from NVIDIA that asked for a minimum of 8 publications at top conferences?

In addition to the problems that exist in academia more broadly, I think many issues in ML can be attributed to the amount of money involved in the industry. How often are fields of study awash in staggering amounts of concentrated wealth without being beset by social dysfunction and perverse incentives?

I sympathize with the young people who want to do ML research just because they're very intellectually curious. They have the misfortune of having to coexist and compete with a larger cohort of people who also have dollar signs in their eyes.

And if you too have dollar signs in your eyes, well this is pretty much what you can expect in any environment that attracts many such people: success is determined to a significant degree by luck and personal connections.

For the most part I ignore conferences and pretty much most of academia when it comes to progress in AI.

Academia and conferences are a joke. Real progress in AI is going to come from industry, and it won’t be published until years after impact, if at all.

I'm a PhD candidate at a UK university but I have 6+ years of prior work experience in the industry as an ML engineer. I have zero intention of staying in the academia and see my PhD as an investment. I've reviewed papers for ECCV, CVPR and NeurIPS to help others as I don't understand how doing blind reviews would further my career in any way. I do my best to read each paper carefully but it's getting insane lately.

I volunteer to review 2-3 papers and get assigned 6 papers, and then 2 more urgent last minute reviews on a Sunday. What I see is academia is full of toxic people, there is almost no one to complaint to and it thrives on making PhD students feel worthless. It works because most PhDs don't have any industry experience and feel like they can't get a job if they quit. I think the solution is to get rid of reviewing on a voluntary basis and stop conferences mooching off from early stage researchers.

The root of the problem is that there are just too many papers and too few best-qualified reviewers. In many other CS fields, papers are reviewed by program committee members who are typically tenure-track faculty. These faculty all know each other, and their personal reputations are on the line when they review. They also know the history of the field and have a broad understanding of what is novel and interesting in a space.

In ML, we have PhD students, and sometimes even undergrads, doing reviewing. They have much less experience than faculty. They also for the most part have no intention of remaining in the academic community, so they have little incentive to build reviewing skills or to build a representation as a reviewer. No wonder the reviews are bad and random.

So the problem isn't really solvably by changing the way the reviewing process works because it's not really a process problem.

High Paying, Highly Innovative, Highly Hyped = Recipe for oversaturstion of students studying it, oversaturation of middle career folks switching their careers into it, recipe for an oversaturated number of non-tech folks completing every LLM, DL certificates to post on their LinkedIn.

What happens with this oversaturation? Your raise the bar to entry - just like what leetcode culture did.

Yay toxic elitism 🤸‍♂️

I was wondering when this topic would inevitably show up after ICLR rebuttal closed. 48 hours it seems is the right amount of time to wait.

Pull up a chair, pour yourself a drink. Let’s commiserate on our collective misery.

The amount of papers that post academia job postings demand from phd students to publish just incentives bad papers and falsifying data.

This completely broken review process is probably the single largest frustration I have with the field. Fundamentally I think the only solution would be to somehow incentivize high quality reviews and potentially punish bad reviewers. Making the identities of the reviewers public afterwards would be one way but I think it creates other problems (such as breeding animosity). My controversial proposal would be to somehow tie your own submissions to the quality of your reviews. Maybe something along the lines of:

• Force the authors of every submitted paper to jointly review something like 3-4 other papers.
• Have meta reviewers who read a given paper and the reviews, scoring the reviews themselves, not the manuscripts. This could be done for some random subset of reviews / manuscripts, not necessarily all.
• Incentivize good reviews, potentially giving a certification of “good reviewer” for accepted papers, displayed publicly, similar to the TMLR certification.
• Punish bad reviewers. Either outright reject their submissions based on their review quality (even if they would get accepted otherwise), or for a less extreme option mark them with a “bad reviewer” certification for their accepted papers as a public badge of shame.

What do people think? Could something along these lines work or is it completely unreasonable?

I fully agree with you OP. But many people, also seniors, realized it for quite some time. Hence, I recommend you to take a look at TMLR. They use OpenReview and the acceptance of your article does not depend on random stuff such as "Impact". Instead, if your claims are correct and somewhat novel, you will get accepted.

Oh, you might enjoy this https://blog.mrtz.org/2014/12/15/the-nips-experiment.html

The problem from a theoretic perspective is that many of the things you recommend might have unintended and in fact, the opposite effect.

• Should reviewers have public statistics tied to their (anonymous) reviewer identity?
  • We do have public statistics tied to our actual profile (our name). You will run into the same reviewers again and again within the peer review process. You'll remember the person a few years ago who was totally unreasonable in the same paper as you. You'll remember that one reviewer who made a brilliant point that the meta-reviewer overrode and turned out they were right. Yes, you'll also run into your former advisors in the peer review process. Try reviewing a paper that your former advisor is a coauthor on and raking it over the coals and finding out later that it was your former advisor's paper. Then some time later your advisor and you are reviewing the same paper side by side, and you have to decide whether you agree or disagree with them.
• Should reviewers have their identities be made public after reviewing?
  • Not sure. This might be good for senior reviewers who do take their job very seriously. But junior reviewers without much experience mess up all the time. Imagine having your social media post from when you were 14 be public, forever, and un-deleteable. That's what it's like to be a junior reviewer and messing up and having it be public. Peer review is much like everything else in academia, an apprenticeship. You learn by doing, and that process requires an element of psychological safety that anonymity can provide.
• Should institutions reward reviewer awards more? After all, being able to review a project well should be a useful skill.
  • I'd love this. We always need more people willing to review well and dispassionately.
• Should institutions focus less on a small handful of top conferences?
  • Institutions do, everyone does. Top conference echo chambers are only for those who think the world revolves around ICML/NeurIPS/ICLR. I'll let you in on a little secret, everyone at COLT laughs at your papers... don't even get me started at what people in STOC think about your papers.

Over and over again I see a paper that is more or less as good as many papers before it, but whether it squeaks in, or gets an oral, or gets rejected, all seem to depend on luck. I have seen bad papers get in with faked data or other real faults because the reviewers were positive and inattentive.

I agree but the problem is also that faked data is incredible hard or even impossible to spot with the current system. You would need to standardize the whole process (code request, exact experiment description, code explanations, creating docker image for reproducability, computational cost, ...). Then the reviewers would need to run some of the experiments themselves aswell (alongside additional experiments to make sure you are not cherrypicking results). This would take a tremendous amount of time and resources

>Should institutions reward reviewer awards more? After all, being able to review a project well should be a useful skill.

While rewarding positive behavior is good, IMO there also need to be negative consequences for bad reviewers. Poor reviews are endemic in the field and there are too many people (often famous people in the field, too) who write shoddy, low-quality reviews and are not punished.

I think one concrete idea is to have ACs (or some other third party) rate reviewer quality, and if someone has low-quality reviews they should be banned from submitting to the conference for a year.

Reviewers are doing this for free on top of everything else in their lives. Shaming them publicly is just going to lead to fewer people reviewing

the review process is completely pointless with regards to reproducibility. the reviewers basically have to go off of what somebody wrote in the paper. other than maybe finding some systematic error from the writeup there is not really much a reviewer can actually detect and criticize (if the model works what else is there to say?). most published papers would be better off as just github projects with a proper descriptive readme that also shows benchmarks anyway. it's not like papers are written very well to begin with. but that doesn't get you a phd.

in physics there is basically no (or minimal) review process and publications are judged by how much your paper got cited. also, there is a full secondary track of researchers who just take other papers and recreate the experiments to actually confirm reproducibility. in ML right now there is no incentive for anyone to just run a published model on different/their own data and confirm that it works correctly. in fact you'd probably be crucified for doing that

On a related note: where is a good place to share a long ML paper (we got told we’re too dense for ICLR, which I agree with) that doesn’t have a 2-year review process (looking at you JMLR)? Subject is tabular synthetic data evaluation.

What's ACs?