We certainly don't want to perpetuate harmful stereotypes. But is it a flaw that the model encodes the world as it really is, statistically, rather than as we would like it to be? By this I mean that there are more light-skinned people in the west than dark, and there are more women nurses than men, which is reflected in the model's training data. If the model only generates images of female nurses, is that a problem to fix, or a correct assessment of the data?
If some particular demographic shows up in 51% of the data but 100% of the model's output shows that one demographic, that does seem like a statistics problem that the model could correct by just picking less likely "next token" predictions.
Also, is it wrong to have localized models? For example, should a model for use in Japan conform to the demographics of Japan, or to that of the world?
If you want the model to understand what a "nurse" actually is, then it shouldn't be associated with female.
If you want the model to understand how the word "nurse" is usually used, without regard for what a "nurse" actually is, then associating it with female is fine.
The issue with a correlative model is that it can easily be self-reinforcing.
I'd say that bias is only an issue if it's unable to respond to additional nuance in the input text. For example, if I ask for a "male nurse" it should be able to generate the less likely combination. Same with other races, hair colors, etc... Trying to generate a model that's "free of correlative relationships" is impossible because the model would never have the infinitely pedantic input text to describe the exact output image.
Randomly pick one.
> Trying to generate a model that's "free of correlative relationships" is impossible because the model would never have the infinitely pedantic input text to describe the exact output image.
Sure, and you can never make a medical procedure 100% safe. Doesn't mean that you don't try to make them safer. You can trim the obvious low hanging fruit though.
You're ignoring that these models are stochastic. If I ask for a nurse and always get an image of a woman in scrubs, then yes, the model exhibits bias. If I get a male nurse half the time, we can say the model is unbiased WRT gender, at least. The same logic applies to CEOs always being old white men, criminals always being Black men, and so on. Stochastic models can output results that when aggregated exhibit a distribution from which we can infer bias or the lack thereof.
This depends on the application. As an example, it would be a problem if it's used as a CV-screening app that's implicitly down-ranking male-applicants to nurse positions, resulting in fewer interviews for them.
Put another way, when we ask for an output optimized for "nursiness", is that not a request for some ur stereotypical nurse?
That's excessively simplified but wouldn't this drop the stereotype and better reflect reality?
I expect that in the practical limit of scale achievable, the regularization pressure inherent to the process of training these models converges to https://en.wikipedia.org/wiki/Minimum_description_length and the correlative relationships become optimized away, leaving mostly true causal relationships inherent to data-generating process.
Perhaps what "nurse" means isn't what "nurse" should mean, but what people mean when they say "nurse" is what "nurse" means.
That’s a distinction without a difference. Meaning is use.
And anyway - contextually -, the representational natures of "use" (instances) and that of "meaning" (definition) are completely different.
For a one-shot generative algorithm you must accept the artist’s biases.
“hey artist, draw me a nurse.”
“Hmm okay, do you want it a guy or girl?”
“Don’t ask me, just draw what I’m saying.”
- Ok, I'll draw you what an average nurse looks like.
- Wait, it's a woman! She wears a nurse blouse and she has a nurse cap.
- Is it bad ?
- No.
- Ok then what's the problem, you asked for something that looked like a nurse but didn't specify anything else ?
Does a bias towards lighter skin represent reality? I was under the impression that Caucasians are a minority globally.
I read the disclaimer as "the model does NOT represent reality".
For example, the most eaten foods globally are maize, rice, wheat, cassava, etc. If it always depicted foods matching the global statistics, it wouldn't be giving most users what they expected from their prompt. American users would usually expect American foods, Japanese users would expect Japanese foods, etc.
> Does a bias towards lighter skin represent reality? I was under the impression that Caucasians are a minority globally.
Caucasians specifically are a global minority, but lighter skinned people are not, depending of course on how dark you consider skin to be "lighter skin". Most of the world's population is in Asia, so I guess a model that was globally statistically accurate would show mostly people from there.
Also, getting a random sample of any demographic would be really hard, so no machine learning project is going to do that. Instead you've got a random sample of some arbitrary dataset that's not directly relevant to any particular purpose.
This is, in essence, a design or artistic problem: the Google researchers have some idea of what they want the statistical properties of their image generator to look like. What it does isn't it. So, artistically, the result doesn't meet their standards, and they're going to fix it.
There is no objective, universal, scientifically correct answer about which fictional images to generate. That doesn't all art is equally good, or that you should just ship anything without looking at quality along various axes.
I want to be clear here, bias can be introduced at many different points. There's dataset bias, model bias, and training bias. Every model is biased. Every dataset is biased.
Yes, the real world is also biased. But I want to make sure that there are ways to resolve this issue. It is terribly difficult, especially in a DL framework (even more so in a generative model), but it is possible to significantly reduce the real world bias.
Sure, I wasn't questioning the bias of the data, I was talking about the bias of the real world and whether we want the model to be "unbiased about bias" i.e. metabiased or not.
Showing nurses equally as men and women is not biased, but it's metabiased, because the real world is biased. Whether metabias is right or not is more interesting than the question of whether bias is wrong because it's more subtle.
Disclaimer: I'm a fucking idiot and I have no idea what I'm talking about so take with a grain of salt.
Yeah, but you get that same effect on every axis, not just the one you're trying to correct. You might get male nurses, but they have green hair and six fingers, because you're sampling from the tail on all axes.
So even if we managed to create a perfect model of representation and inclusion, people could still use it to generate extremely offensive images with little effort. I think people see that as profoundly dangerous. Restricting the ability to be creative seems to be a new frontier of censorship.
Do they see it as dangerous? Or just offensive?
I can understand why people wouldn’t want a tool they have created to be used to generate disturbing, offensive or disgusting imagery. But I don’t really see how doing that would be dangerous.
In fact, I wonder if this sort of technology could reduce the harm caused by people with an interest in disgusting images, because no one needs to be harmed for a realistic image to be created. I am creeping myself out with this line of thinking, but it seems like one potential beneficial - albeit disturbing - outcome.
> Restricting the ability to be creative seems to be a new frontier of censorship.
I agree this is a new frontier, but it’s not censorship to withhold your own work. I also don’t really think this involves much creativity. I suppose coming up with prompts involves a modicum of creativity, but the real creator here is the model, it seems to me.
Interesting idea, but is there any evidence that e.g. consuming disturbing images makes people less likely to act out on disturbing urges? Far from catharsis, I'd imagine consumption of such material to increase one's appetite and likelihood of fulfilling their desires in real life rather than to decrease it.
I suppose it might be hard to measure.
I won't speak to whether something is "offensive", but I think that having underlying biases in image-classification or generation has very worrying secondary effects, especially given that organizations like law enforcement want to do things like facial recognition. It's not a perfect analogue, but I could easily see some company pitch a sketch-artist-replacement service that generated images based on someone's description. The potential for having inherent bias present in that makes that kind of thing worrying, especially since the people in charge of buying it are likely to care, or notice, about the caveats.
It does feel like a little bit of a stretch, but at the same time we've also seen such things happen with image classification systems.
Propaganda can be extremely dangerous. Limiting or discouraging the use of powerful new tools for unsavory purposes such as creating deliberately biased depictions for propaganda purposes is only prudent. Ultimately it will probably require filtering of the prompts being used in much the same way that Google filters search queries.
If the model only generated images of female nurses, then it is not representative of the real world, because male nurses exist and they deserve to not be erased. The training data is the proximate causes here, but one wonders what process ended up distorting "most nurses are female" into "nearly all nurse photos are of female nurses" something amplified a real world imbalance into a dataset that exhibited more bias than the real world, and then training the AI bakes that bias into an algorithm (that may end up further reinforcing the bias in the real world depending on the use-cases).
