An Open Letter to the Communications of the ACM

Flamingo:

(1)(2) I think we're agreed in part. The industry hires on merit and that's been my experience both sides of the desk. You claim something about altruism, I didn't. You claim something about industry motives I didn't*.

Correct me if I'm wrong, but I think you are then implying that therefore I think all is well and nothings to be done, but that is not true. I think tech, if used right could revolutionize and economically scale education starting from best practices parenting and individually tuned education all the way through. I think that's the foundation stone in rebuilding our infrastructure. But it's a completely different topic.

(3) We disagree. IMHO, you train the best performing models and then should state a transparent policy on why you change those results. This policy may be "too little data for that class" on up to "that's hate speech".

(4)(5) They are neither wimps nor stupid. You are right, struggle sessions often resulted in murder, "cancel/gang on/trial by internet" mostly aim at career, but tastes the same.

* In my experience ~ high up in a derivatives operation and at Intel, then 6 startups, 3 I founded and my non-profit. At Intel, and Willow Garage hiring women and minorities was prioritized. In startups, we were resource constrained. My/our motive: hire the best people whom I/we could afford. Period. IMO, this is the common case. Many women and minorities were hired as a result.

@Bradski, I will keep it to the point.

(1), (2) Asian-American families have a high median income (i.e., they have access to good schools), so the participation of local born Asian-Americans in CS is not surprising but cannot be credited to the system. Foreign-born Asians get their undergraduate degrees in their home countries (their foundations already laid) and US tech industry is fortunate to tap into top talent without much investment (with some paying 0% tax!) . The point is the industry employs Asians because it helps their bottomline. There is nothing altruistic about it. If you want to know their representation in a globalized field, you will have to denominate it using the global population.

(3) The proof is in the pudding! Whether you tweak the model or filter its outputs, it does not matter to the overall outcome. So, this is subjective and is a design choice ("how", not "what).

(4) So, what you are saying is that your mentees are not "wimps" but they confided in you that they are afraid of signing this letter? See the contradiction? "I don't suffer wimps" is not a factual point in a rational debate. I just happen to think that most rational individuals act in their self-interest and do not express their opinion freely when there is even an iota of chance that their opinions may affect their success. This is why we have double-blind reviews. So, when you ask your mentees of their opinion, they know your opinion on this matter, so they will likely tell you what you want to hear.

(5) I have read about "struggle sessions" but calling a rigorous intellectual debate with colleagues on Twitter cannot be compared to the tactics of communist China.

Flamingo:

To clarify, when (1) when I said:

"'over represented' in the field in the US" I meant "'over represented' in the field in the US".

(2) You have accused me of proposing that AI researchers are "like a stream of bat's piss"! To which I say: Um, Ahem ... they "shine out like a shaft of gold when all around is dark"! 'course I never said nor ever implied dat pissy straw man.

(3) Since 2 got twisty, I'm twisting back, except twisting Hume: If your network is learning "Is" and you figure out how to insert some "oughts" into the nodes. Well, that's anti-Hume-an! He warned ya. Now humans are a mess of tribalistic hates, greeds, lusts for power, kindness, madness, loves, wants, needs and shockingly different behaviors, standards and results. That comes out in the data and none look great. I think it would be a rich area to take Pearl further so that a network could learn some differential causal model of behavior and results about it all. But for now, output the Is as near as it can be and then deal with it in the sunlight. One real issue is a catch-22: Your behavior is bad, so I rate you bad so you never have a chance. Yeah, serious policy issue there that you and I moral people want to rectify at more levels than just the model outputs.

(4) Your supposition is wrong, I don't suffer wimps. The people I champion because they and I know they're good.

(5) If they made a weak argument, I'd have DM'd their are and said: "Crikey! Dat's weak friend, amp your game!" That didn't happen. But, do read up on struggle sessions. I shan't accuse you of supporting them, because it's not true.

Responding to 👇👇👇

tl;dr; (1) Asians (people from East Asia and India) are NOT over-represented if you normalize their share to the global population, which is the right way to measure. (2) The success of Asians against all odds is NOT a proof that prejudice/bias against non-whites does NOT exist. This is cherrypicking or using an exception to further a flawed generalization. (3) Adding a filter on the model outputs to effect policy decisions is NO MORE moral or immoral than fixing the model. (4) Women and minority researchers who (you say) agree with your opinion are subject to the same pressures (i.e., they may be telling you what you want to hear), so your rebuttal suffers from the same fallacy of the argument that you are trying to rebut! (5) Most often the "meritorious elites" are NOT canceled, it's just that they end up making a weak argument and leave Twitter on their own when pointed out. Of course, there is more pressure on them because they are held to a higher standard. But these "cancel culture" fake tears are often made after losing a debate.

Expanded response:

Science is a very globalized endeavor. So, when you want to make an argument as to who is under-represented or over-represented, you gotta consider the relative of groups wrt their global population. China+India constitute 30% of world population. So, they are not "way over represented" as you falsely assert. But yes, they are reasonably well-represented than people from sub-Saharan Africa. Nevertheless, while western European whites had structural advantages built by years of colonialism, slavery etc, Asian success is a fight against all odds. So, their numbers cannot prove that prejudice does not exist. In fact, the US immigration policy favored western Europeans over others until the 60s (on top of historical advantages from colonialism, slavery, racism that favors lighter skin, etc).

I don't quite understand why adding filters at the model output is anymore kosher than changing the model to fix the inherent bias in it. In fact, fixing the model seems to be the most direct approach to addressing the root of the problem. Also, your toy example about loans is flawed because you assume that resources are unlimited. If you have a resource limit, we have to make a decision between granting a loan to a minority person (whose chances are repayment are calculated based on some aggregate group statistics, not based on his personal history alone) for buying his first house versus a rich, white guy (who happens to find favor with the model) who wants to buy his third vacation house. The root cause of this problem lies in the model which associates an individual to a group and extrapolates the group's aggregate statistics to determine whether he will qualify or not. It is fundamentally against the idea of free will. Any and all biases are unnatural, against free will, and not objective at all.

The last part of women and minority researchers privately agreeing with you does not prove anything. They might have agreed not to get into your wrong side and are subject to the same pressure you allude to them not signing this letter. Also, most "meritorious" people face pressure in a medium like Twitter because they are held to a higher standard, they have more followers (so they get more responses to their tweets), and they sometimes cannot take the fact that they were wrong. When your argument is aggressively rebutted, it's NOT cancel culture. It's free speech working as intended.

I do remember the trial by twitter that got Yann LeCun to quit twitter because he correctly stated that the a GAN that turned Obama's face into a white man* was doing so because it was trained on a particular database that contained more Caucasian. I defended him as simply being correct and was also correct and so I just left too since like Maoist struggle sessions of old, the more you try to reason in such cases, the more it is taken as "proof" that you are non-woke. I also noted that struggle sessions often targeted meritorious elites because the attackers gain more social power in doing so and seeing on twitter the glee that he left twitter. I believe it was (a) bad behavior and (b) wrong.

(b) above is a big category spanning moral wrong, but also wrong as in an incorrect causal factor keeping minorities from the field** and it's also wrong to train AI models that don't accurately represent the biases in the data. Doing so means you've intentionally made the model wrong in a hidden way. You may indeed want to apply a policy or filter on the resulting outputs via a dynamically adjusted set of filter data so that we directly know what is being filtered. For instance, in applying for a house loan, the model should indicate chance of repayment, period. If some poor, possibly minorities then have difficulty getting loans, this should be addressed by social policy/subsidy/support, but not by suppressing qualified people just to round out some numbers which would be bald face and morally abhorrent profiling.

Maybe as Pressler suggested, it might reference a list of incidents, but we did have someone seemingly w/o consequence listing people to cancel in the field. I've served as reference to a number of woman and minority researchers and asked them what they think of this letter. The consensus so far is that they agree with the letter but would never sign as it would hurt their career and also some feel that their technical work seems a slower road to recognition than the political work of others.

* Note that the GAN also often turned him into a Black man and also noting that Obama is indeed half "White".

** Noting that people of India or East Asian descent are way "over represented" in the field in the US but somehow don't count. The asserted cause of prejudice seems wrong and so the solution won't work for anything but hypocrisy.

Fine sentiments, and important for protecting junior researchers and other vulnerable individuals in science.

One thing, though - the letter also clearly condemns Pedro's behavior against Anima, when he called on the mob to contact her employer and report "evidence" against her for the NeurIPS leadership. Does he now disavow that behavior? What's good for the goose ought to be good for the gander.

The problem is that those of us who are at higher risk (junior researchers, minimal branding, controversial pasts, depend on favors from everyone) are not going to be able to sign it unless anonymously, but being anonymous carries little weight.

The problem is that those of us who are at higher risk (junior researchers, minimal branding, controversial pasts, depend on favors from everyone) are not going to be able to sign it unless anonymously, but being anonymous carries little weight.

It sends a very weak message indeed when one of the major proponents of this letter, Pedro Domingos, personally attempts to "cancel" leading female researchers in AI. The most notable recent attempt was his decision to personally reach out to Nvidia, CalTech, and NeurIPs conference organizers this month to "curtail..." Dr. Anima Anandkumar's "disgraceful behavior". In his own words, "...assemble evidence... Justice is coming."

Perhaps the other supporters are signing this letter in good faith, but as another commenter noted, no evidence has been provided to support the fundamental claims, and no obvious benefit is suggested resulting from the ACM reaffirming its current principles.

Pedro Domingos is a massive hypocrite who happily uses his seniority to harrass people, attack and silence them based on their identity and views. PS What do you think about calling a fellow researcher "deranged"?

Where were y'all when Steven Salaita's offer was rescinded??

This letter, ostensibly aimed at focusing research and engineering on the “very nature of scientific inquiry” seems to fail at exactly the problems it condemns. The letter is vague rather than precise, invites generalisations rather than establish applicability, and makes explicit or implicit empirical assertions without empirical support. In other words, it is the opposite of what we’d expect from STEM professionals.

Let’s examine it in some detail:

> writing with sadness and alarm about the increasing use of repressive actions aimed at limiting the free and unfettered conduct of scientific research and debate.

This clause makes two claims: that there is an increasing use of “repressive actions” and that their aim is to limit “the free and unfettered conduct of scientific research and debate”. These are empirical claims that can and should be established in a letter signed by these people and purportedly advancing these goals — perhaps not with the same rigour as mathematical proof or physical experimentation, but surely it can, should, and must be supported with at least the rigour we’d like to see in the social sciences?

The first claim requires some time-series data. The second speaks to motive, which is harder to establish, but it could be more objectively changed to speak to outcome, i.e. that the result of those actions is to limit free scientific research. That would also require some time-series data showing a marked decline in the volume or variety of research, and some way to show that it is, indeed, caused by the said repressive actions.

If these two claims cannot be established, then surely we, as STEM professionals, must assume they are not true, regardless of personal feeling, observation, and political perspective any one of us may have on the matter.

> We condemn all attempts to coerce scientific activities into supporting or opposing specific  social-political beliefs, values, and attitudes, including attempts at preventing researchers from exploring questions of their choice, or at restricting the free discussion and debate of issues related to scientific research.

I would assume that the signatories also condemn all attempts at mass genocide perpetrated by invading aliens, but the implication here is that the condemnation is directed at a non-empty, and, moreover, a non-negligibly-small set, yet this has not been established. It’s insufficient that someone might *believe* some action is aimed at “coerc[ing] scientific activities into supporting or opposing specific social-political beliefs”, but it’s required to objectively establish that this is, indeed the goal, or, at the very least, the result (i.e. there is a measurable decline in the variety or volume of research and that it is caused by such actions).

This is especially important because similar claims made in other disciplines, when examined, were found to be without merit [1]. It is somewhat paradoxical, and perhaps cause for embarrassment, that a letter calling for “[s]cientific work [to] be judged on the basis of scientific merit, independent of the researcher's identity or personal views,” makes scientifically examinable claims without the requisite merit, which invites the question, if unsupported by merit, and without even recognising the need to support claim with scientific enquiry, could it be that this letter is precisely an example of what it condemns?

[1]: https://journals.sagepub.com/eprint/UC8HG8URH2WQWVIWN5AG/full, https://youtu.be/8Ax_dliY2hM