Read on Twitter
"didn’t meet our bar for publication ... It ignored too much relevant research".
He says of our paper, with over 128 cross-discipline citations. And so fires one of the authors. 2020's monster version of Reviewer #2. https://twitter.com/JeffDean/status/1334953632719011840
He says of our paper, with over 128 cross-discipline citations. And so fires one of the authors. 2020's monster version of Reviewer #2. https://twitter.com/JeffDean/status/1334953632719011840
Let's take a look at the average number of citations for one of @JeffDean 's favorite conferences, NeurIPS.
A rousing 25 citations is the max.
https://www.microsoft.com/en-us/research/project/academic/articles/neurips-conference-analytics/
https://www.microsoft.com/en-us/research/project/academic/articles/neurips-conference-analytics/
Let's check out ICML, another leading ML conference.
https://www.microsoft.com/en-us/research/project/academic/articles/icml-conference-analytics/
WOAH almost 27 whole citations!
https://www.microsoft.com/en-us/research/project/academic/articles/icml-conference-analytics/
WOAH almost 27 whole citations!
Let's take a look at which citations are "missing" to such an extent that the paper is unpublishable and unfixable within 2 months.
1) "it talked about the environmental impact of large models, but disregarded subsequent research showing much greater efficiencies"
Nerp--no disregard. The paper is about harms and risks.
Nerp--no disregard. The paper is about harms and risks.
One cool thing about being a researcher is that you get to learn a LOT about background contexts. It is by using that context that you can isolate and hone in on what your unique research contributions should be.
For example, say everyone is really into technology X. But once in awhile, there are concerns about technology X being bad in some way. In order to develop technology X responsibly, we need to critically examine what the concerns of that technology are.
Ironically, I had a great conversation with Jacob Devlin (BERT guy) about the paper. He thought it was cool, immediately started thinking through how you could train a model *not* to learn stereotypes, and we grooved on the role of supervision in such a model.
That's the kind of fun and forward looking discussion we need to be having in AI right now.
2) "it raised concerns about bias in language models, but didn’t take into account recent research to mitigate these issues."
The authors have worked on bias mitigation, so not taking into account their own domains of research would be quite literally impossible.
The authors have worked on bias mitigation, so not taking into account their own domains of research would be quite literally impossible.
The question, then, is of the *specific biases* discussed in the paper, which are the ones that have had great advances in mitigation strategies?
For example, say we're talking about the linguistic characteristics of sexism.
What's the mitigation strategy for that?
I must have missed that paper.
What's the mitigation strategy for that?
I must have missed that paper.
"A cross functional team then reviewed the paper"...so, not a cross functional team of those qualified to review the paper.
"as part of our regular process"
A process being "regular" doesn't mean it's not a process that disproportionately censors a minority population.
It is precisely these "regular" systems that we need to rethink and rebuild in order to create a new culture in AI development.
A process being "regular" doesn't mean it's not a process that disproportionately censors a minority population.
It is precisely these "regular" systems that we need to rethink and rebuild in order to create a new culture in AI development.
Maybe even a culture where it's possible to retain a Black woman for more than 3 years!
"Highlighting risks without pointing out methods for researchers and developers to understand and mitigate those risks misses the mark on helping with these problems. "
I'm glad @JeffDean actually knows the right answer for how to develop ethical AI!
I'm glad @JeffDean actually knows the right answer for how to develop ethical AI!
And here I was, thinking I had some reasonable research view worth discussing on the approach. d'oh!
But's he's right. A paper focusing on harms and risks should be required to be split in half to be half harm, half benefit. All the other papers only highlighting benefits are fine though.
A relevant thread from @emilymbender touches on the importance of both defining problems and solving problems in AI research. https://twitter.com/emilymbender/status/1333418175380144129
"Our aim is to rival peer-reviewed journals in terms of the rigor and thoughtfulness in how we review research before publication. " --> Did this earlier say 'to be even more rigorous than peer-reviewed....' ? Either way, let's talk about peer review and blind review.
In 2001, double-blind review was introduced by the journal Behavioral Ecology...There was a significant increase in female first-authored papers, a pattern not observed in a very similar journal that provides reviewers with author information. https://pubmed.ncbi.nlm.nih.gov/17963996/
A 2012 study in the Proceedings of the National Academy of Sciences showed that identical application materials for a lab position were more favorably read when the name at the top of the resume was “John,” rather than “Jennifer.” https://www.pnas.org/content/109/41/16474
Double-blind review helps us to mitigate some of the effects of knowing an author's race, gender, or reputation, and instead focus in on the content of the paper.
I believe that I am obligated to speak up when my employer publicly belittles me and/or my colleagues' scholarship.
Broadcasting a misleading or biased narrative to promote a particular political view "misses the mark" on doing good research.
