Yann LeCun quits Twitter amid acrimonious exchanges on AI bias

If someone is making a good faith rational argument, and the response is "I find your conclusion potentially offensive, so I will suppress your expression"; that is the suppressor acting in bad faith.

The energy around "doing what feels right despite evidence" and "the ends justify the means" have turned into the destruction of nuance. It's very hard to say "I agree with your conclusion, but not your reasoning" because how one arrived at a conclusion no longer matters, only that the conclusion fits the dogma.

It is anti intellectualism.

It's single-minded, irrational, damaging to society and acts with little regard for the nuance of actual reality - which means it cannot in the long-term solve actual problems we face without creating bigger ones.

Acting is good faith and with intellectual honesty is always going to require voluntary participation. But if one contestant decides to knock the pieces off the table, then what we're watching is not a chess game, but one party claiming chess superiority because he bit his opponent on the neck.

Think about that.

For the sake of ethical R&D, it's counter-productive to build a hierarchy of investment into the problem. Admittedly, the responsibility of end-results can differ, but the consensus that this ethical work is important should ideally be universal. That said, there should not be some ivory tower where you wash your hands of the ethical problems of your field.

Recently seems in support of a black scholar who cried racism because her non peer-reviewed work that was only posted on arxiv wasn't cited in a lecture on GANs ...

Voicing these opinions would probably label me as racist in their book ironically.

Why not just edit the end results to show what you really want?

And that’s the deeper problem here: “it’s just a biased dataset” is a misdiagnoses. It’s a whole system of biases that leads to people thinking they are training with balanced data when they manifestly are not.

You’re never really going to achieve this mythical “balanced training data” until you untangle all of the other implicit personal and organizational biases. There are a whole host of ethical discussions that need to happen to even begin to flesh out what “balanced” might even mean for, say, facial recognition software intended for use in law-enforcement, but the same biases that lead people to skip right past those discussions and begin training are often the very ones that result in the biased data to begin with.

It could also be balanced so that the evaluation metrics were similar for each subgroup (possibly ending up with a sample that's very different from the population). But what are the subgroups? For any commonly used definition of race, there is a lot of intra-group variety.

Maybe balancing the training data is enough. But figuring out what balance even means is a huge question.

I don't think there's a ready CS/stats solution to this problem, so it will require interdisciplinary engagement and listening to the people who have been on the wrong end of facial-recognition bias is likely a place to start.

For example, there are already attempts to use expert systems in the courts - to "scientifically" determine flight risk (and thus bail), and even chance of re-offence (and thus punishments). Those are trained on datasets of convictions. But we know that both conviction rates, and arrest rates, are themselves indicative of bias - e.g. some people get arrested more often for a crime, for which other people might get a verbal warning from police.

So ML takes those biases, and entrenches them further. Worse yet, because people tend to believe that "computers cannot be biased", the fact that it's ML that's producing a specific result, is taken as prima facie evidence that said result is immune to criticism from that perspective. And with most advanced models, you can't really crack it open and make sense of what's inside - so you can't easily prove that it is biased.

For example imagine some ML algorithm that predicts the potential value of a customer (or their likelihood to steal from you) as they walk in the door of your business. How do you remove the racial bias from a system like that when racial biases exist in society that result in the average White person being in a higher socioeconomic class than the average Black person? The ML algorithm will tell you White customers are more valuable than Black customers. It is doing exactly what you told it to do, but the end result is that Black people get worse service in your business.

If you're looking at things that correlate, you've got a lot of choices. Saying that your data set is representative, your algorithms are correct, doesn't mean your choices of what correlations to use are objectively or morally correct.

Suppose that you have a system that accurately tells you men have higher car insurance claims on average than women.

The same system might also be able to tell you people with low credit scores have higher claims on average than people with high scores.

Is it right or ok to use the first correlation and ignore the second? Or vice versa? Or maybe both of them are unacceptable? There's nothing objectively inherent in the correlations that tells you it's ok to charge men who are good credit risks the same as men who are bad credit risks. Or optimal economically! Those are independent and difficult questions.

Imagine you have two groups of people that have different means and distributions of “input” (could be intelligence, wealth, interests, ...), and this input is somehow correlated with some output (insurance rate, criminality, university acceptance rate). Both input and output are one-dimensional.

Then you only have two choices. (1) You “unify” the distribution and use it as joint input, resulting in one group (the one with the higher mean) getting “more” - many people think this is unfair. (2) Or you can subtract the means first and so make the distributions as equal as possible, which can result in a person from the “more” group having worse outcome than an equivalent person from the “less” group (if the means are close together and the distributions are wide, there will be many such people) - many people would consider this unfair.

There are plenty of examples, current and historical, of society picking one or the other, for different groups and topics: wealth/success/income (old vs young, Jews&Asians vs other Americans, men vs women, whites vs blacks), university acceptance (Jews used to be discriminated against (look up “Jewish problems”), now Asians are, men vs women [1], “positive” discrimination of blacks) - just a few. In many examples historically, we consider (1) fundamentally better than (2) - e.g. discrimination against Jews at universities - but it seems it takes the society a long time to reach this conclusion in each instance.

[1] https://en.wikipedia.org/wiki/Simpson%27s_paradox#UC_Berkele...

When more variables come up and get more complicated it seems like it would be impossible to say "You're based on race" versus "You're based on a multitude of factors, several of which correlate to race, and consequently you give more or better loans to race X over Y."

Yeah this is tough, answering the question of the odds a person may default on a loan will probably reflect historical bias. In fact I'd argue that it should show this if the model is accurate.

The moral issue is what do you do with that information? Do you use that knowledge to improve the outcomes of the disenfranchised or do you reinforce the bias based on the information you have?

No decision is completely safe and no system gets everything right in the first pass.

All we can ask for is a serious effort to understand risks sufficiently well to reasonably believe that the system is safe.

Have government grants for ML research be contingent on training on diverse datasets. Fund research into more diverse benchmarks and push them as the goal to beat.

If they are in here I hope they don't misconstrue my linking this image. I think they explained themselves very well and I feel bad that they were thrust into the middle o f this controversy.

My point is that while diversity might help, you need a lot more than that to address this problem.

To be clear no one (or at least no one of note) has their hackles raised over this specific dataset. PULSE is fine for what it is, and no one criticized PULSE for having these results.

It is however a great demonstration for laypeople about how ML models aren't magic and don't always do what you, as a human, would expect. This is true irrespective of the source of that unexpected behavior.

That said, I believe the disclaimer you mention was added only after the recent twitter discussion.

Philosophically I find ML’s tendency to reflect the biases we bring to it very revealing. In some ways it shows us what we’ve built, the underlying biases that we’d rather argue about and ignore. When an algorithm selects longer sentences for black men than white men, we rightly see that as racism. Some say, “use better data, we’ll then be objective!” But I wonder if a better initial reaction is, “wow, look how badly out system has failed that it would produce such a bad dataset.” Never mind that maybe it’s not actually possible to be objective and that’s the point. Math doesn’t lie, maybe when we make a racist model we’re failing to see the mirror it’s holding up for us.

Hello, mister/lady, our 100% unbiased system has automatically determined that you have commited XYZ crime, would seem more appropriate.

ML tools are developed, and filter down into the general developer population where they are used without fully comprehending the biases they contain or can contain if used incorrectly. I'm pretty partial to the idea that ML is a systemic bias footgun. VERY hard to not use incorrectly, and with potentially HUGE social repercussions. E.g. the youtube funnel towards extremism that was well documented a few years ago - visitors start on some innocuous video, and get recommended more and more extreme things, funnelling traffic towards the alt-right in an unhealthy way.

More likely, the reason that bias isn't routinely fixed is that it isn't easy, and these kinds of biases do make it into production systems. Which makes it a net positive in my view that the occasional shitstorm reminds society of this fact.

Can that be unpleasant for AI researchers? Sure... but if it bothers then, then perhaps they could focus their research on trying to fix the problem?

Physicists had and have unpleasant conversations about their moral responsibility for nuclear weapons. Other fields of research should take their moral responsibilities seriously as well (not just AI research, by the way).

I also think there's value in simple true statements.

I'm not saying she's wrong, I'm saying we don't know because she defaulted to the argument that "you should listen to minorities", not "here is the evidence".

What's more, every single example of injustice in her tutorial was an image recognition/classification problem - entirely different from the generative model that originally sparked the debate.

Should cryptography researchers backdoor their own papers, because terrorists or pedophiles might use it?

This case is so pervasive that it gets taught in basically every university that teaches ML - and yet those students go on to repeat the exact problem they were educated against.

Assume you want to train an AI to recognize shops or buildings for example for a Google car.

Well if you do it in the US with skyscraper, in Europe, in Africa, Middle-East or Asia, you will get completely different results and biases.

Also, I don't see how anyone has the resources to compensate such a social bias, unless they plan to do a shooting trip in hard to access locations and try to justify such as "Seriously officer, the reason I take all those photos is to make sure my machine learning model is unbiaised so I don't classify shops in your country as shacks."

And if we forget human activity, even looking at nature, the fauna and flora are different between areas their color, how sparse they are, etc.

Last example, applying sparsity to human activity, what is actually a town in some country might be classified as a settlement for example due to bias.

It has real world consequences that can drastically damage communities.

Abstracting away the ethical issues of your (elite) employment is a personal choice that is anything but objectively neutral.

No, this is only technically correct but actually wrong. In the example, they did in fact fail to build the AI, as is their job. "Recognize white faces" is a lame research goal, "recognize human faces" is the real thing. So if somebody builds an AI system that fails on out-of-sample data, then says that had they tried to do it properly they would have succeeded, that's a pretty lame excuse for a poor AI system. They didn't even do their job in the narrow sense that you're using, you don't even need to consider the "social biases" or whatever, it's just a system that didn't work. In fact, many years into this research program, "focussing on his task and correcting it" (working on all human faces) is still not done given the current performance of these systems, but they are quite sure it can be done if they tried.

1) The model uses biased data when it could have used unbiased data.

2) The model uses biased data when no unbiased data exists and is incapable of correcting for these discrepancies.

The first case is most clearly and engineering/implementation issue, the second is obviously not. Biased data is a known failure case of ML, its the responsibility of researchers to design for it.

I think he has some responsibility to at least acknowledge the complexity of this issue in such cases. Not speaking to the public is also a option if he doesn't like the leader role, so him deleting twitter is a completely ok thing in my eyes.

"'Once the rockets are up, who cares where they come down? That's not my department' says Wernher von Braun"

By the way, ML is not nearly as new as you seem to think, and given the amount of resources poured into it by FAANG-type companies recently, even five years is a lot of time for ML nowadays.

Because we're all responsible for how the tools we built are used and what they enable, and how they affect society at large. That's what ethics is about, something which seems to be absent in the education of the modern citizen and in particular engineer or scientist, who is supposed to come to work, program things and not think too much about the impact their products have on the world.

Gebru’s CVPR 2020 talk Computer vision in practice: who is benefiting and who is being harmed? again addressed the role of bias in AI, “I think that now a lot of people have understood that we need to have more diverse datasets, but unfortunately I felt like that’s kind of where the understanding has stopped. It’s like ‘let’s diversify our datasets. And that’s kind of ethics and fairness, right?’ But you can’t ignore social and structural problems.“

"""

Seems that these two paragraphs alone describe what she has been saying. Where did your "algorithmic bias" assertion came from? Who mentioned algorithmic problems?

> If there is any sort of fairness, ethics or what have you initiative @Facebook and @facebookai, let me tell you how little credibility they have because you have the loudest microphone out there.

Its interesting how one perceives this as "throwing shit" when it is harsh criticism directed at what she denotes as the little power any ethics or fairness org at FB has.

The other quotes are out of frustration and are in the civil territory. If those are uncivil to you and the others, well, that's your line of sensitivity as it is definitely not the bar for incivility. If something sounds harsh, it doesn't mean it is uncivil. It is very easy to twist words or remove them out of the conversation context to make them seem like fighting words.

I could be very harsh in address people here and still be civil about it. The air I smell here 'It is Yann LeCunn, legendary figure, never criticize what he says or if you do so, do it very very very politely even if you have to do it 1k times.'

I'm no longer interested in finding out more perceptions on this. I saw what I saw and you guys see what you see. I don't see any uncivil activity just sensationalism at an opportune time and a mob of anti-social justice trying to attack whatever seems "woke." These issues would have to be dealt with no matter what period of time.

I've had people call my behavior "unprofessional" in the past because they had a very narrow set of expectations for "professional." Whatever that seemed to fit them per person. Always glad to have a third party to sort it out with them. I regard unprofessional as a potential dogwhistle these days since the only people who understand what it really means are the ones who knows it is being used in a narrow subjective manner. They are just loaded terms.

I have no idea who this person is, but it sounds like he’s not very good at engaging with the public on Twitter, so the best strategy for him is to leave the platform.

No one disagrees with that statement. The point that the ethicists are making is that that isn't the whole story, and that focusing only on the lowest hanging fruit (dataset bias) is ultimately harmful since it shifts focus away from the other, deeper problems.

1. Responded with a detailed summary of his views.

2. Said he was open to changing his mind in the face of convincing arguments; indicated a desire to learn.

3. Asked people to stop attacking his opponent, in response to unfair attacks on her.

4. Only a week ago, sent multiple tweets out about the problems with a paper which demonstrated racial bias. https://medium.com/@CoalitionForCriticalTechnology/abolish-t...

There's something in between not taking someone seriously and agreeing with their argument.

Paper: https://arxiv.org/pdf/2003.03808v1.pdf

Here are the authors: Sachit Menon, Alexandru Damian, Shijia Hu, Nikhil Ravi, Cynthia Rudin

They conveniently captured an image of them all in one place for us here - https://cdn.telanganatoday.com/wp-content/uploads/2020/06/Au...

Pretty diverse team if you ask me. Sprinkling underrepresented (and therefore 'underpowered') representation into engineering teams will solve a lot of issues, but I think it has yet to be proven that it's going to help ML bias.

Suggested technical solutions [1 above] was exactly one sentence and the only sentence in the presentation that suggested any sort of technical approach, presented without elaboration, and the recommended solutions like Counterfactual Testing are also methods for fixing the training data.

The remaining 99% of the presentation was spent on the idea that if AI engineers and scientists were more diverse and attended more diversity training, the models they produce hopefully wouldn't be biased.

> If you have diverse datasets, but your ... model doesn't like ... things correlated with being black... what then?

How do you release your data/models today? Does your release process include a section detailing what the data implies or the limitations of the mode? How would your boss handle being exposed to these findings? If your boss still wants to use the data set or deploy the model knowing the limitations, do you care enough to do something about it? Maybe you have a different set of answers to these questions. Does that make asking them any less important?

It would indeed be a valuable contribution if Gebru could have shown how better training data would not have fixed the problem (e.g. by feeding better training data to the same model and still reproducing the issue).

But as it is, she is asserting that better training data won't help without proof and not recommending any alternatives, while taking a hostile tone of conversation. I would be hard pressed to take her claims as good faith criticism.

- https://www.propublica.org/article/machine-bias-risk-assessm...

- https://themarkup.org/locked-out/2020/05/28/access-denied-fa...

- https://ainowinstitute.org/AI_Now_2019_Report.pdf

I can recommend more if you get through those.

This isn't allowed anymore, for better or worse. If you don't talk about things with a social justice lens on Twitter, you're liable to get called out as ignorant. It's the #1 subject on the site right now, and no place for discussing anything remotely adjacent to it without invoking it.

Personally, I don't think that an interest in the social impacts of one's work as a researcher should be considered hype -- shouldn't one be interested in their impact on the world? For computer vision in particular, these systems are widely deployed. Also, they are just part of a larger apparatus of automatic decision systems that everyone now interacts with every single day, from getting your insurance rates to getting a loan to... Those have now become prosaic too -- certainly not scifi.

If you steer clear of certain topics[1], this place is absolutely civil.

[1]: Not the usual suspects and not what divides the rest of the world...

i mean, i can give plenty others that do exactly what you said. twitter is not unique and certainly not the problem.

i read his final(?) tweets and the reasons why he decided to leave twitter. and i don't think or remember him blaming the medium.