AI researcher Timnit Gebru resigns from Google • Hacker News Board

Re: AI researcher Timnit Gebru resigns from Google

#871

Post by bmitc » Fri, Dec 04, 2020, 9:02 PM UTC

Footkerchief wrote: post #845

mehrdada wrote: post #752

Earlier quoted context omitted.

> This seems more of a case of selective enforcement, where she had been allowed to go around the process in the past (or felt that she has enough leverage to be able to do so now) and then not being allowed to do so this time. Jeff's email explicitly says the paper was "approved" then submitted, albeit perhaps in one day. The two-week thing might be a policy but would be immaterial if it ended up approved anyway wit…

I agree -- the passive voice here (intentionally?) makes it impossible to tell whether the external submission was done in good faith: > Unfortunately, this particular paper was only shared with a day’s notice before its deadline — we require two weeks for this sort of review — and then instead of awaiting reviewer feedback, it was approved for submission and submitted. If (to the best of her knowledge) she got appro…

I found the quoted sentence to be very unclear, which is surprising because it is the crux of the matter. From it, I can't tell the timeline between the order of when things happened and who did what. Who approved and submitted what and when?

Re: AI researcher Timnit Gebru resigns from Google

#872

Post by nebulous1 » Fri, Dec 04, 2020, 9:06 PM UTC

cmiles74 wrote: post #821

sjs382 wrote: post #397

It's hard to make too much sense about what happened. Here's what I made of it, if anyone can point out any important mistakes/inaccuracies/details: 1. She and her co-authors submitted a paper for review 1 day before its (external, for publication?) due date. It's supposed to be submitted to Google 2 weeks prior, so they can review. 2. The authors submitted it externally before the review was returned. 3. Google deci…

I'm curious about this "submitted paper for review 1 day before publication" assertion. From my reading, it sounded like work on this paper had been going on for a while. The email quoted on "Platformer" makes it sound like communication with HR (which seems weird, honestly, for a research paper) had been going on for at least two months.[0] > Imagine this: You’ve sent a paper for feedback to 30+ researchers, you’re…

I don't think google are saying they weren't aware of the paper prior to the day-before, just that they hadn't seen the actual paper itself before then.

Out of curiosity do people think that a company either should or would publish self-funded research that's detrimental to their PR? I'm not sure that this is a realistic expectation with our current methods of company governance (which should probably be changed). I'm assuming the research isn't overly safety oriented and is more like "overuse of google leads to decreased overall well-being" rather that "use of the cigarettes that we sell leads to significantly reduced life expectancy". Of course that statement begs the question of where the line between the two lies.

Re: AI researcher Timnit Gebru resigns from Google

#873

Post by sudosysgen » Fri, Dec 04, 2020, 9:07 PM UTC

chokeartist wrote: post #835

sudosysgen wrote: post #416

Earlier quoted context omitted.

I'm not sure I understood your comment correctly, so I'll apologize in advance if I misunderstood. That statement is contained in a paper I linked that dates from 1980. I'm pretty sure it's factually correct, because I'm trying to make nothing but a demonstration that an idée reçue, that ML systems are only based due to their data, is incorrect. There's really no cloaked message behind it, and I'm not sure why one wo…

> That statement is contained in a paper I linked that dates from 1980. I'm pretty sure it's factually correct, because I'm trying to make nothing but a demonstration that an idée reçue, that ML systems are only based due to their data, is incorrect. That's reasonable. I may have taken the initial text in the wrong context. Thanks for clarifying. To expand on my prior comment of "Fuck x group", perhaps the tone of pr…

I think there is a bit of a double standard here. I try in general to be as charitable to both sides of an agreement possible. If your standard for uncharitability is such that the linked tweet is a proof of racism, I don't understand why one would take a tweet claiming against overwhelming evidence at face value.

Personally, I don't think either indicate racism. I see the link you tweeted as an excessively abrasive response at a dismissive and reductive understanding of an issue, bias in learning system, despite research old and new. At the same time, I see LeCun's dismissal as a genuinely held position in error.

To see the first as racist but the second as okay seems to be a case of being much more charitable to one side than the other, though I understand how it may look this way.

Re: AI researcher Timnit Gebru resigns from Google

#874

Post by cmiles74 » Fri, Dec 04, 2020, 9:29 PM UTC

nebulous1 wrote: post #872

cmiles74 wrote: post #821

Earlier quoted context omitted.

I'm curious about this "submitted paper for review 1 day before publication" assertion. From my reading, it sounded like work on this paper had been going on for a while. The email quoted on "Platformer" makes it sound like communication with HR (which seems weird, honestly, for a research paper) had been going on for at least two months.[0] > Imagine this: You’ve sent a paper for feedback to 30+ researchers, you’re…

I don't think google are saying they weren't aware of the paper prior to the day-before, just that they hadn't seen the actual paper itself before then. Out of curiosity do people think that a company either should or would publish self-funded research that's detrimental to their PR? I'm not sure that this is a realistic expectation with our current methods of company governance (which should probably be changed). I'…

I did find a Reddit post with the abstract and some discussion from a person who claims to be a reviewer.

https://www.reddit.com/r/MachineLearning/comments/k69eq0/n_t...

In my opinion, it seemed pretty low key. But I've only read the abstract.

In terms of Google's expectations and what should we consider reasonable, I'm really not sure... I do think that when they hired this person to be "technical co-lead of the Ethical Artificial Intelligence Team at Google", they probably should've guessed that there could, possibly, be a conflict between "make loads of money" and "ethical".

Re: AI researcher Timnit Gebru resigns from Google

#875

Post by dmix » Fri, Dec 04, 2020, 9:40 PM UTC

tolbish wrote: post #655

dmix wrote: post #541

Earlier quoted context omitted.

No, James was just a random engineer who got fired after the VP of Diversity took issue with the memo he shared. Firing people for sharing unpopular opinions is a whole other nest of bees. One that I'm sure Timmit and her ilk would fully support. So he wasn't exactly the person I had in mind. Nor was Timmit fired and I don't believe Damore was pushing unconditional demands. But that said I can see some similarities w…

I thought he was fired for insinuating that his female colleagues were hired for reasons other than their merit, not just for sharing a memo with unpopular opinions.

I'm still not sure what any of that has to do with my original comment, he was fired for "breaching the code of conduct". Other than maybe(?) the comment was some attempt at a "but they do it too" sort of reply you find on Reddit.

Judging by your interpretation of James memo I'm leaning towards that direction. Especially considering the fact James made multiple direct claims in the memo he was not suggesting anything of the sort which you are alleging (I have a feeling few people actually read the memo).

Regardless the two stories were very very different. Nor does his story match my comment's 'archetype' of a certain type of employee who always end up flaming out of their job, usually via making non-conditional demands (unsurprisingly coming from a bubble/worldview where it's entirely normal to have things which are not open to discussion).

Re: AI researcher Timnit Gebru resigns from Google

#876

Post by ska » Fri, Dec 04, 2020, 9:57 PM UTC

fantod wrote: post #593

ska wrote: post #516

Earlier quoted context omitted.

That sounds like something easily rectified, so why push for retracting the submission?

I think the logic is that the paper would need to be retracted, rectified, and then resubmitted. Since the submission deadline was already passed, resubmission would be impossible.

Submissions deadline is typically for evaluation, not camera ready. It's not unusual for textual changes to make their way in after that point, either in response to reviewer or to improve the text in a way that doesn't fundamentally change the results.

Re: AI researcher Timnit Gebru resigns from Google

#877

Post by dheera » Fri, Dec 04, 2020, 10:23 PM UTC

johbjo wrote: post #836

dheera wrote: post #711

Earlier quoted context omitted.

That sounds like a horrible use of human hours, especially at Google salary level. Paper writing is painstaking and mentally tiring. It would be much more efficient if they reviewed results and an abstract and then decided based on that "Go ahead and write the paper and publish it" or "Don't waste time writing it up".

Money is not an issue at all. It's about protecting the perception of quality that comes from having Google co-authors, and ultimately the Google brand. Otherwise, everyone and their mother would milk the Google name by submitting papers in every direction.

Sure, then they should have a 2-step process. The first step should be approving writing the paper, and any high selectivity should happen at that point.

The second step should be approving the final contents, and that process should not reject a high fraction of submissions and should allow resubmissions with changes.

It would suck to spend 200 hours figuring out how to fight to position figures in LaTeX only to be rejected by some dudes who think the figure is confidential.

Re: AI researcher Timnit Gebru resigns from Google

#878

Post by rndgermandude » Fri, Dec 04, 2020, 10:25 PM UTC

joshuamorton wrote: post #708

rndgermandude wrote: post #706

Earlier quoted context omitted.

I lost the assumption of good faith already in her first tweet, in particular >You can’t just reduce harms caused by ML to dataset bias. She's already attacking a strawman right there. Yann did not deny any harms caused by ML.

Nor did Timnit claim that he was. Her disagreement was about the causes of harms, not the existence of resulting harms.

There was no disagreement. Yann didn't say anybody about harms nor did the guy he was replying to (who talked of dangers). In particular Yann did not suggest in any way a) there are no harms or b) that harms are only related to biased training sets. Yann was commenting on the outcome of a particular research project and how they used a biased training set resulting in the outcome that was observed.

Timnit brought up harms first, then pretended Yann did marginalize such harms and attributed them solely to a biased training sets. And then viciously attacked that strawman. That's a bad faith argument.

I can appreciate that she might have been indeed generally sick and tired as she writes, and can appreciate that sick and tired people will not always manage to be nice or overcome their own biases and assume good faith all the time from the other party; we're all human after all. But that doesn't change anything about her argument being made in bad faith.

Re: AI researcher Timnit Gebru resigns from Google

#879

Post by ianbicking » Fri, Dec 04, 2020, 10:32 PM UTC

woko wrote: post #846

ianbicking wrote: post #596

Earlier quoted context omitted.

> And she was under the impression that her company would value her so much that, not only they wouldn't take the first "clean" opportunity to let her go, but that they'd want her to stay so much that she could leverage it? In a realpolitik sense you are probably right. But right now I'm going through standard onboarding training and they are VERY explicit that it's not supposed to work this way, and that you are not…

You are right, but off-topic in this specific context. Indeed, it was neither retaliation, nor exclusion. She offered her employer a "clean" opportunity to let her go, and they accepted it. It will be difficult for her lawyers to prove that she was let go as a retaliation, which it is not. The only thing which is apparent is that her employer did not value her work as much as she thought, so that her demands and her…

There is "legally liable for retaliation" and "morally liable". I think you are right, they aren't legally liable. I personally try not to be morally liable for an injustice, even if my actions are legally defensible, and I expect the same from the people I work with. I think many people at Google feel similarly, which is why the question of legal liability will not settle the matter internally.

But yeah, another lesson from this: never offer to quit.

Re: AI researcher Timnit Gebru resigns from Google

#880

Post by joshuamorton » Fri, Dec 04, 2020, 10:33 PM UTC

rndgermandude wrote: post #878

joshuamorton wrote: post #708

Earlier quoted context omitted.

Nor did Timnit claim that he was. Her disagreement was about the causes of harms, not the existence of resulting harms.

There was no disagreement. Yann didn't say anybody about harms nor did the guy he was replying to (who talked of dangers). In particular Yann did not suggest in any way a) there are no harms or b) that harms are only related to biased training sets. Yann was commenting on the outcome of a particular research project and how they used a biased training set resulting in the outcome that was observed. Timnit brought up…

> Yann didn't say anybody about harms nor did the guy he was replying to (who talked of dangers)

This feels like unreasonable semantics. The dangers are precisely the danger of causing harm. The harms therefore are concrete results of theoretical dangers manifesting. They aren't different.

> In particular Yann did not suggest in any way a) there are no harms

I agree, and I've said as much.

> b) that harms are only related to biased training sets

He did, insofar as he suggested that the dangers were due solely to bias in the training set, which is implied when he says that if you train the same model on different data, everyone looks African. Like yes, that is true, but it doesn't reduce the harms (or dangers, if you want to be precise). It just creates a different set of biases with a different set of potential harms (which again, are "dangers").

I'm not seeing a strawperson.