[chomp]

Insensitive to anyone who has a moderate amount of understanding of machine learning and social empathy.

You can't plug your ears and say "it's just your training set" as a response to unfairness in ML algorithms. Real life is biased. Any real life data in our world is going to be biased. If you train algorithms on this data, they will cement any existing divides in society. So, with the understanding that researchers need to be more circumspect about ML algorithms than worrying about just the training data, consider that the upsampling algorithm in question only worked for white people because they fed it a huge amount of white faces. Claiming "it's just the training data" is one of those "well yes, but actually no" situations where ML researchers tend to miss the broader picture of how ML algorithms are used in real life, and just makes Yann look ignorant.

[devalgo]

The real argument they were making against Lecun is whether a mathematical function can be biased. Care to explain how a gradient is racist?

[chomp]

>Care to explain how a gradient is racist?

Sure. Your comment's language equivalent is something along the lines of "Care to explain how words are racist?" Which yes, they are just a collection of words. They possess no consciousness and cannot be racist by themselves.

Similarly, a gradient is just a collection of vectors. It's just numbers. However, like language, it's what they represent that matters.

For example, I can create a machine learning algorithm to determine who should get a home loan. I create a gradient to optimize the algorithm to give loans to people who I think are unqualified.

The gradient can easily be racist if it optimizes heavily on something like race. Minorities tend to be lower income and so can be seen as less qualified as higher income individuals. However that's the easy argument, and also quite illegal. If you exclude race, there's 2nd degree variables that are proxies for race. Things like zip codes, job titles, whether they rent or buy. These are not explicitly illegal to filter on, though the end result is illegal if they exclude certain protected statuses. It can even be no fault of the researchers who implement the algorithm, because controlling for bias using real world data is extremely difficult. But we must do it, since it is the ethical thing to do.

And so, it's easy to see that one can optimize ML algorithms to exclude certain protected statues, which is what can make the algorithms racist.

[devalgo]

You failed the test. The Gradient is not biased, the data is. This was of course LeCun's point... This is pure foolishness

[chomp]

Maybe I'm not explaining it very well. Look, so things have meaning deeper than their face value. To use a really basic example, The number 14 means nothing, it's a number. The number 88 means nothing. In the same context, they mean something not good.

There are English words that as pieces, they don't mean anything except their face value. I can string words together that mean bad things that are harmful to real humans.

Gradients are not racist by themselves, they're just math. It's like saying multiplication is racist.

But I can use multiplication as a tool in a chain to create weighted averages to create a naive Bayesean classifier to reject people for home loans.

And so too can I misapply gradient descent as a part of a larger ML model that is racially biased. For instance, I could choose a loss function that when minimized, gives biased output despite less biased input. Or, I could accidentally settle on a local minimum on the gradient in my model. There's many naive implementations of an algorithm that will just be biased no matter the unbiased inputs.

So in summary, a gradient is just math and is not racist by itself. It's being used in an algorithmic tool chain that researchers are frequently using which potentially will always produce biased output no matter the inputs (but more often than not also with biased input).

[kevingadd]

It should be self-evident that if you add race as a variable the resulting function at the very least could easily end up racist. If you add biases to a function it will be biased. Which is fine, sometimes the biases are necessary to solve difficult problems!

Even if you insist that a gradient or mathematical function is unbiased and can never have negative impact based on race or gender or other demographics, you have to explain any resulting negative impact somehow. Saying that the function or gradient is racially biased is a generous interpretation of the situation because it allows the creators to deflect blame towards an error in their mathematics or training set. If you insist on claiming that the training set and mathematics are infallible, one of the only remaining explanations is that the creator intended to discriminate. I'd rather not assume that!

[devalgo]

All you have done is make the case that the data was biased. A mathematical function is not racist.

[twsttest]

I'm not disagreeing with your basic idea, but it seems you're nitpicking and talking past Yann's point.

A model's only link to the real world is the training data, so saying it's sufficient to "worry about the training data" captures all the concerns we may have about bias, because from the model's POV there is no other relevant interface with the real world.

Saying "we need to do more" is devoid of meaning when by addressing the training data we are truly doing all we can as model builders and trainers.

[chomp]

So here's an example of more that we can do.

A huge problem in the field is that we must use the previous benchmarks. This is because how do you know if the needle moves or not if you just change your data constantly?

So. In order to tackle this problem, someone with more resources than me needs to create training sets that are less biased. THEN, new academic papers need to benchmarked against the old biased sets, and also the new "less biased" (I don't think it's possible to ever get 0% bias, the world just isn't that clean) sets. And progress needs to be eventually transitioned to be measured on the new less biased sets.

The upsampling algorithm used pictures of celebrities. And the researchers put a blurb in their paper that was basically a "We know this is biased but everyone uses it so we must also". I feel like this is less useful science than an algorithm trained on more of a mix of actual real-world humans.

I admit it's quite challenging and probably impossible to do in some areas. I mean, how do you make a field whose end algorithmic goal is generalization, not use real world data to generalize people? But I think the issue can be worked on, and the need to use celebrity photos to train a set is a good place to start.

[newen]

All this is going to do is researchers not releasing data and code when publishing their articles so that the public doesn't meme biases/mistakes of their data/code into twitter hate mobs.

We'll probably go back to the 2000s model where you have to email the authors for code and data. The authors will delay by saying they are preparing it and then release it a few years later when it becomes irrelevant for public discourse.

[chomp]

ML is a huge field outside of modelling humans and their behavior. For instance, image recognition of vehicles, financial data prediction and analytics, and weather forecasting, to name a couple examples. Those don't draw scrutiny. The problem comes with generalizing humans. And generalizing using biased data. And applying generalized algorithms in areas that cause a lot of harm. I think these researchers should properly be placed under the microscope since they have the potential to be very hurtful to society. I do not think they should be subject to death threats or loss of income or whatever the social media mob throws at them these days, but I don't think researchers should be cavalier in creating algorithms that generalize humans without taking very careful steps to not create bias in the end result.

[scoutt]

Sorry, but was the trained ML model to be implemented and used, as is, in public, like in an airport? Or was it to become the next standard or the next "ML for dummies" book? Or was it just research or an experiment?

If it was an experiment, then let it be. Perhaps the researcher was looking for something else, circumscribing the data, model, whatever to the experiment itself.

> researchers need to be more circumspect about ML algorithms

What does entitle you to tell what to study or how?

[chomp]

Your entire comment is correct, but still missing the bigger picture. It's understood that it's way easier to detect features in pictures of white faces than black faces due to the fact that it's easier to resolve lines and shadows. These lighting differences show up once the image is pixelated, and gives something for PULSE to lock on to when it attempts the upscale. I'm questioning whether or not the algorithm even works for cases where these lighting differences are difficult or impossible to resolve.

If the researchers created a toy, then great, it's a cool project and is a neat algorithm. But they didn't create a toy. It's an academic paper to attempt to move the needle forward in ML academia. And they are doing the exact same thing as a lot of other researchers, which is basing their research on old biased benchmarks. If the bedrock of the field is based on biased data and everyone builds on top of that, your research down the line will skew more and more in favor of the bias.

>What does entitle you to tell what to study or how?

Nothing entitles me. It is my opinion based on the facts in front of me. The ML field has a bias problem, researchers toss a "oh this is biased" blurb in their papers, and then continue using the biased data. Everyone looks at the cool demos, and then the research gets slurped up and implemented without regard to the science. More algorithms get based on previous biased algorithms.

[scoutt]

> doing the exact same thing as a lot of other researchers, which is basing their research on old biased benchmarks

They might have a reason. I can understand if they want to compare the result of the model with a past experiment. That's normal.

> attempt to move the needle forward

Completely agree, so just let them work.

By the way, I don't see "evil" in these experiments and I want a 100% free from bias model too, but I wouldn't dare to attribute the result to lazyness, stupidity or racism. If I come with something completely new then I would try to compare it with something that already exists too.