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.
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.
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.
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".
Put another way, when we ask for an output optimized for "nursiness", is that not a request for some ur stereotypical nurse?
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.
That’s a distinction without a difference. Meaning is use.
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.
What percent of people should be rendered as white people with broccoli hair? What if you request green people. Or broccoli haired people. Or white broccoli haired people? Or broccoli haired nazis?
It gets hard with these conditional probabilities
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.
Your description is closer to how the open source CLIP+GAN models did it - if you ask for “tree” it starts growing the picture towards treeness until it’s all averagely tree-y rather than being “a picture of a single tree”.
It would be nice if asking for N samples got a diversity of traits you didn’t explicitly ask for. OpenAI seems to solve this by not letting you see it generate humans at all…
And anyway - contextually -, the representational natures of "use" (instances) and that of "meaning" (definition) are completely different.
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.
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.
That's excessively simplified but wouldn't this drop the stereotype and better reflect reality?
> We investigated sex differences in 473,260 adolescents’ aspirations to work in things-oriented (e.g., mechanic), people-oriented (e.g., nurse), and STEM (e.g., mathematician) careers across 80 countries and economic regions using the 2018 Programme for International Student Assessment (PISA). We analyzed student career aspirations in combination with student achievement in mathematics, reading, and science, as well as parental occupations and family wealth. In each country and region, more boys than girls aspired to a things-oriented or STEM occupation and more girls than boys to a people-oriented occupation. These sex differences were larger in countries with a higher level of women's empowerment. We explain this counter-intuitive finding through the indirect effect of wealth. Women's empowerment is associated with relatively high levels of national wealth and this wealth allows more students to aspire to occupations they are intrinsically interested in.
Source: https://psyarxiv.com/zhvre/ (HN discussion: https://news.ycombinator.com/item?id=29040132)
For a one-shot generative algorithm you must accept the artist’s biases.
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.
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.
How does the model back out the "certain people would like to pretend it's a fair coin toss that a randomly selected nurse is male or female" feature?
It won't be in any representative training set, so you're back to fishing for stock photos on getty rather than generating things.
“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 ?
I have a feeling that we need to be real with ourselves and solve problems and not paper over them. I feel like people generally expect search engines to tell them what's really there instead of what people wish were there. And if the engines do that, people can get agitated!
I'd almost say that hurt feelings are prerequisite for real change, hard though that may be.
These are all really interesting questions brought up by this technology, thanks for your thoughts. Disclaimer, I'm a fucking idiot with no idea what I'm talking about.
Preliminarily and provisionally. Then, they start discussing their concepts - it is the very definition of Intelligence.
Are the logical divisions you make in your mind really indicative of anything other than your arbitrary personal preferences?
This is a far cry from say the USA where that would instantly trigger a response since until the 1960s there was a widespread race based segregation.
Which real world? The population you sample from is going to make a big difference. Do you expect it to reflect your day to day life in your own city? Own country? The entire world? Results will vary significantly.
If it did, would you believe that’s a real representative nurse because an image model gave it to you?
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.
1. The model provides a reflection of reality, as politically inconvenient and hurtful as it may be.
2. The model provides an intentionally obfuscated version with either random traits or non correlative traits.
3. The model refuses to answer.
Which of these is ideal to you?
If I ask for pictures of Japanese people, I'm not shocked when all the results are of Japanese people. If I asked for "criminals in the United States" and all the results are black people, that should concern me, not because the data set is biased but because the real world is biased and we should do something about that. The difference is that I know what set I'm asking for a sample from, and I can react accordingly.
Your logic seems to rest on this assumption which I don't think is justified. "Skewing search results" is not the same as "hiding the biases of the real world". Showing the most statistically likely result is not the same as showing the world how it truly is.
A generic nurse is statistically going to be female most of the time. However, a model that returns every nurse as female is not showing the real world as it is. It is exaggerating and reinforcing the bias of the real world. It inherently requires a more advanced model to actually represent the real world. I think it is reasonable for the creators to avoid sharing models known to not be smart enough to avoid exaggerating real world biases.
A model that returns 100% of nurses as female might be statistically more accurate than a model that returns 50% of nurses as female, but it is still not an accurate reflection of the real world. I agree that the model shouldn't return a male nurse 50% of the time. Yet an accurate model needs to be able to occasionally return a male nurse without being directly prompted for a "male nurse". Anything else would also be inaccurate.
Well the results would unquestionably be biased. All results being black people wouldn't reflect reality at all, and hurting feelings to enact change seems like a poor justification for incorrect results.
> I'd say it doesn't actually matter, as long as the population sampled is made clear to the user.
Ok, and let's say I ask for "criminals in Cheyenne Wyoming" and it doesn't know the answer to that, should it just do its best to answer? Seem risky if people are going to get fired up about it and act on this to get "real change".
That seems like a good parallel to what we're talking about here, since it's very unlikely that crime statistics were fed into this image generating model.
curiously, this search actually only returns white people for me on GIS
Every model will have some random biases. Some of those random biases will undesirably exaggerate the real world. Every model will undesirably exaggerate something. Therefore no model should be shared.
Your goal is nice, but impractical?
I said "It is reasonable... to avoid sharing models". That is an acknowledged that the creators are acting reasonably. It does not imply anything as extreme as "no model should be shared". The only way to get from A to B there is for you to assume that I think there is only one reasonable response and every other possible reaction is unreasonable. Doesn't that seem like a silly assumption?
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.
"It is possible for a man to provide care" is not the same statement as "it is possible for a sexually dimorphic species in a competitive, capitalistic society (...add more qualifications here) to develop a male-dominated caretaking role"
You're just asserting that you could imagine male nurses without creating a logical contradiction, unlike e.g. circles that have corners. That doesn't mean nursing could be a male-dominated industry under current constraints.
—
The AI doesn’t know what’s common or not. You don’t know if it’s going to be correct unless you’ve tested it. Just assuming whatever it comes out with is right is going to work as well as asking a psychic for your future.
If they were to weight the training data so that there were an equal number of male and female nurses, then it may well produce male and female nurses with equal probability, but it would also learn an incorrect understanding of the world.
That is quite distinct from weighting the data so that it has a greater correspondence to reality. For example, if Africa is not represented well then weighting training data from Africa more strongly is justifiable.
The point is, it’s not a good thing for us to intentionally teach AIs a world that is idealized and false.
As these AIs work their way into our lives it is essential that they reproduce the world in all of its grit and imperfections, lest we start to disassociate from reality.
Chinese media (or insert your favorite unfree regime) also presents China as a utopia.
No it is not, because you don’t know if it’s been shown each one of its samples the same number of times, or if it overweighted some of its samples more than others. There’s normal reasons both of these would happen.
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.
“When I use a word,’ Humpty Dumpty said in rather a scornful tone, ‘it means just what I choose it to mean — neither more nor less.’
’The question is,’ said Alice, ‘whether you can make words mean so many different things.’
’The question is,’ said Humpty Dumpty, ‘which is to be master — that’s all.”If the only way to do AI is to encode racism etc, then we shouldn't be doing AI at all.
We may not like what it shows us, but blindfolding ourselves is not the solution to that problem.
Mental definition is that "«artificial»" (out of the internal processing) construct made of relations that reconstructs a meaning. Such ontology is logical - "this is that". (It would not be made of memories, which are processed, deconstructed.)
Concepts are internally refined: their "implicit" definition (a posterior reading of the corresponding mental low-level) is refined.
Perhaps what "nurse" means isn't what "nurse" should mean, but what people mean when they say "nurse" is what "nurse" means.
So? Draw your consequences.
Following what was said, you are stating that "a staggering large number of people are unintelligent". Well, ok, that was noted. Scolio: if unintelligent, they should refrain from expressing judgement (you are really stating their non-judgement), why all the actual expression? If unintelligent actors, they are liabilities, why this overwhelming employment in the job market?
Thing is, as unintelligent as you depict them quantitatively, the internal processing that constitutes intelligence proceeds in many even when scarce, even when choked by some counterproductive bad formation - processing is the natural functioning. And then, the right Paretian side will "do the job" that the vast remainder will not do, and process notions actively (more, "encouragingly" - the process is importantly unconscious, many low-level layers are) and proficiently.
And the very Paretian prospect will reveal, there will be a number of shallow takes, largely shared, on some idea, and other intensively more refined takes, more rare, on the same idea. That shows you a distinction between "use" and the asymptotic approximation to meanings as achieved by intellectual application.
Is it? I'm reminded of the Microsoft Tay experiment, were they attempted to train an AI by letting Twitter users interact with it.
The result was a non-viable mess that nobody liked.
I say this because I’ve been visiting a number of childcare centres over the past few days and I still have yet to see a single male teacher.
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).
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.
Given that male nurses exist (and though less common, certainly aren't rare), why has the model apparently seen so few?
There actually is a fairly simple explanation: because the images it has seen labelled "nurse" are more likely from stock photography sites rather than photos of actual nurses, and stock photography is often stereotypical rather than typical.
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.