μ-site

Triaging Boring Publications

This short essay is a response to Daniel Lemire’s post. Originally, I intended for this to be quick tweet, but after line 35 of my response, I realized that it’s too long for Twitter and needs to be an essay on its own.

Tl;dr

Interesting read; I do have some reservations, though. Generally speaking, I see the same symptoms, but think the diagnosis is too narrow to be generally applicable. Instead of the mono-causal need for “customers,” I’d attest a melange of a variety of shortcomings, which in combination lead to slow pace:

Opposing view: Need for customers

I am opposing the view that science needs to have a customer. There are practical fields, where this clearly is useful and reasonable. But there are also fields that are at odds with this requirement. Many subjects have theoretical fields, where there may not be a customer available right now. A customer may emerge in 100 or 200 years in the future. To mandate a customer could effectively stifle future research endeavors.

Multi-causal vs mono-causal

Fashion

Over the past ten years, I’ve often felt that parts of computer science I am familiar with—programming languages and computer security, both to a varying degree—have been subjected to fashions. Among the non-CS scientists I know, the feeling is pretty much the same. The fashion aspect in science is pervasive: it influences choice of topics, relative ranking of topics, being “first,” importance of venues, and certainly a lot of other things. The key downside of this fashion circuit is that it often moves to fast for science to matter and/or mature: When one gets past the low-hanging fruit and the fashion circuit moved elsewhere, what are you supposed to do?

Conferences

Computer-science conferences need reconsideration, too. I’ve always held the opinion that computer science is not special and that the conference-focus is merely an artifact of its age. Conferences are also subject to fashion, but the major issue—in my opinion—is that the function of conferences seems to be defunct. A programme committee (PC) should accept papers in an objective, unbiased, and informed fashion.

When you take 100pct of the papers, probably half of which that are not a good fit (premature submission, wrong venue, subpar language, etc.) can be identified relatively quickly. But what about the rest? Depending on the acceptance rate, some papers get in, others don’t. The acceptance rate has become a proxy for selectivity and prestigiousness, but in reality it is often random and more of a social process of the PC members. (A couple of studies seem to confirm the random nature of PC decisions, cf. NIPS experiment.) The paper selection, thus, is not necessarily objective.

Another big downside of conferences is PC selection, which is also affected by social and sometimes also institutional processes and biases. Large research centers often have substantial funding, but only for a given duration of a couple of years, after which they will be evaluated. Having multiple members of one such research center serve on a single programme committee could result in back channeling. I have not heard about any such case, but it is clear that such a possibility exists and the only way to ensure that it doesn’t is to prevent such PC configurations. The paper selection, thus, is not necessarily unbiased.

From an individual perspective, reviewers often tend to comfortably rank themselves as experts, even when they only have passing familiarity in the actual field of expertise. I feel this is a clear case of Hanlon’s razor, since I have not seen anybody doing it maliciously. Since reviewers are not accountable for writing bad reviews, there is simply no downside to writing bad reviews, and neither to consider oneself an expert when it is clearly not the case. The paper selection, thus, is not necessarily informed.

Goodhart’s law

Goodhart’s law is also complicit. Scientists are smart people and will figure out a way to game the system. If the number of publications is relevant, they will figure out way to maximize publication count. Replace publications with any kind of goal, then that will be optimized for.

Perils of systems research

Michael Stonebraker summarized his fears for database research a couple of years ago, and I feel his diagnosis holds for systems research in general. I talked to several senior members of the PL community and they, too, confirmed that PLDI before the 00s was a lot more about ideas without a solid, industrial-strength implementation and full-blown evaluation. Nowadays, a grad student has to implement something for two years to obtain interesting results, only to be shot down by a reviewer in what often appears to amount to less then a couple of work hours by a reviewer.

Another downside of systems research is that the community is often acting in a self-destructive way. Some people prefer to reject a great idea demonstrated in a sub-par implementation. Others argue that implementation in framework A is stupid, and the authors should “in good taste” use framework B. There are infinite ways of constructing such arguments, but the truth of the matter is: people tend to lose sight of the forrest for the trees. A good idea with a shoddy implementation is much better than a shoddy idea with a good implementation. Insecure and overblown egos are damaging the field, alienating junior members and stifling systems research.

What are we to do?

We need an environment conducive and committed to research. Personally, I believe that we need to stop the nonsensical and gigantic waste of tax payer money to maintain the conference circuit that shows diminishing returns. If we had journals, with smaller focus and smaller, but actual expert reviewers, we could ensure progress in all subfields regardless of the fashion du jour. By having two or three journals in a sub-field, one could focus on early-stage/prototypical vs late-stage/finished work. Goodhart’s law ensures that this switch is not ideal and will also have downsides.

The alternative is to maintain the status quo. Whether this is wise or not, I do not know but I have serous doubts; certainly, the choice is not up to me. I want to end on a positive note by quoting Einstein’s definition of insanity: “Doing the same thing over and over again and expecting different results.

(c) stefan brunthaler, 2021-01-02, 23:17