it's a hard problem. at least they tried.
1. The training data would've been the best way to get organic results, the input is where it'd be necessary to have representative samples of populations.
2. If the reason the model needs to be manipulated to include more "diversity" is that there wasn't enough "diversity" in the training set then its likely the results will be lower quality
3. People should be free to manipulate the results how they wish, a base model without arbitrary manipulations of "diversity" would be the best starting point to allow users to get the appropriate results
4. A "diverse" group of people depends on a variety of different circumstances, if their method of increasing it is as naive as some of the are claiming this could result in absurdities when generating historical images or images relating to specific locations/cultures where things will be LESS representative
If I want to produce "animal images" but it only produces images of black cats, do you think there is any question whether it's a problem or not?
The "issue" is a different one: that training data - IE, reality, has _unwanted_ biases in it, because reality is biased.
Producing images of men when prompting for "trash collecting workers" should not be much of a surprise: 99% of garbage collection/refuse is handled by men. I doubt most will consider this a "problem," because of one's own bias, nobody cares about women being represented for a "shitty" job.
But ask for picture of CEOs, and then act surprised when most images are of white men? Only outrage, when proportionally, CEO's are, on average, white men.
The "problem" arises when we use these tools to make decisions and further affect society - it has the obvious issue of further entrenching stereotypical associations.
This is not that. Asking DALLE for a bunch of football players, would expectedly produce a huddled group of black men. No issue, because the NFL are disproportionately black men. No outrage, either.
Asking DALLE for a group of criminals, likewise, produces a group of black men. Outage! Except statistically, this is not a surprise, as a disproportionate amount of criminals are black men.
The "problem" is with reality being used as training data. The "problem" is with our reality, not the tooling.
Except in the cases where these toolings are being used to affect society - the obvious example being insurance ML algorithms. et al - we should strive to fix the issues present in reality, not hide them with handicapped training data, and malformed inputs.
I think, for about 95% of the world football is synonymous with soccer. Its kind of interesting that you take this particular example to represent what reality looks like statistically
This is not great. Only about 57% of NFL players are black, and the percentage is more like 47% among college players. It would be better to at least reflect the diversity of the field, even if you don't think it should be widened in the name of dispelling stereotypes.
> Asking DALLE for a group of criminals, likewise, produces a group of black men. Outage! Except statistically, this is not a surprise, as a disproportionate amount of criminals are black men.
Only about 1/3 of US prisoners are black. (Not quite the same as "criminals" but of course we don't always know who is committing crimes, only who is charged or convicted.) That's disproportionate to their population, but it's not even close to a majority. If DALLE were to exclusively or primarily return images of black men for "criminals", then it would be reinforcing a harmful stereotype that does not reflect reality.
"criminals" producing most black people actually would be a perfect example of bias in DALL-E that is arguably racism.
Black people commit a diproportionate amoumt of crime (for a variety of socioeconomic reasons I won't get into here), but even so white people make up a majority of criminals (because white people are the largest ethnic group by far).
Thus, a random group of criminals, if representive of reality, should be majority white.
A mortgage AI that calculates premiums for the public shouldn't bias against people with historically black names, for example.
This problem is harder to tackle because it is difficult to expose and resign the "latent space" that results in these biases; it's difficult to massage the ML algo's to identify and remove the pathways that result in this bias.
It's simply much easier to allow the robot to be bias/racist/reflective of "reality" (its training data), and add a filter / band-aid on top; which is what they've attempted.
when this is appropriate is the more cultured question; I don't think we should attempt to band-aid these models, but for more socially-critical things, it is definitely appropriate.
It's naive on either extreme: do we reject reality, and substitute or own? Or do we call our substitute reality, and hope the zeitgeist follows?
That's a great example, thanks. Also, I hope the teams working on that come up with a different solution...
My question to you is: is an algorithm that takes no racial inputs (name, race, address, etc) yet still produces disproportionate results biased or racist? I say no.
[1] https://files.stlouisfed.org/files/htdocs/publications/es/08...
The government, and many people, have moved the definition and goal posts; so that anything that has the end result of a non-proportional uniformity can be labeled and treated as bias.
Ultimately it is a nuanced game; is discriminating against certain clothing or hair-styles racist? Of course. Yet, neither of those are explicitly tied to one's skin color or ethnicity, but are an indirect associative trait because of culture.
In America, we have intentionally muddled the waters of demarcation between culture and race, and are starting to see the cost of that.
Not that I agree with that but I don't see why you would build one otherwise, if you wanted discrimination free mortgages wouldn't the whole process by anonymized and minimal personal information rather than the current system of having to hand over every detail of your life.
It's a hard problem for sure. But remember, the bias ends with the user using the tool. If I want a black scientist, I can just say "black scientist".
Let me be mindful of the bias, until we have a generally intelligent system that can actually do it. I'm generally intelligent too, you know.
That is a really, really, narrow viewpoint. I think what people would prefer is that if you query "Scientist" that the images returned are as likely to be any combination of gender and race. It's not that a group is "fragile", it's that they have to specify race and gender at all, when that specificity is not part of the intention. It seems that they recognize that querying "Scientist" will predominantly skew a certain way, and they're trying in some way to unskew.
Or, perhaps, you'd rather that the query be really, really specific? like: "an adult human of any gender and any race and skin color dressed in a laboratory coat...", but I would much rather just say "a scientist" and have the system recognize that anyone can be a scientist.
And then if I need to be specific, then I would be happy to say "a black-haired scientist"
It's way bigger than just this narrow race issue the current zeitgeist is concerned about.
But I agree, maybe I should skew to being optimistic that at least we're trying