4 ms·

Several comments don't seem to understand what the "unfair bias" mentioned is. It doesn't have anything to do with censoring your favorite conservative search r

Several comments don't seem to understand what the "unfair bias" mentioned is. It doesn't have anything to do with censoring your favorite conservative search result. The machine learning "bias", at least the low hanging fruit, is learning things like "doctor == male", or "black face = gorilla". How fair is it that facial recognition or photo algorithms are trained on datasets of white faces or not tested for adversarial images that harm black people? Or if you use translation tools and your daughter translates careers like scientist, engineer, doctor, et al and all of the pronouns come out male? The point is that if you train AI on datasets from the real world, you can end up reinforcing existing discrimination local to your own culture. I don't know why trying to alleviate this problem triggers some people.

minikites 8y ago
"Machine learning is like money laundering for bias" https://twitter.com/pinboard/status/744595961217835008 https://twitter.com/pinboard/status/744595961217835008
zawerf 8y ago
It's definitely a problem worth alleviating. But it is "triggering" because it is an open problem to determine whether a bias is harmful, even for human beings. So it becomes an impossible/unreasonable amount of extra work if you demand it as a prerequisite. For example, in your translation tool example, even a human translator would have trouble making the least offensive translation possible. She/he/(insert favorite pronoun here) would need to realize the audience is a young impressionable child who is about to base her entire world-view on whether there's statistically more of her gender in that one sentence of translation. For a machine learning algorithm to understand enough about human nature to not offend the parent of that child, you're better off waiting for AGI that can surpass human tact.
- cromwellian 8y ago
  You don't think a black person trying to use a photo management app and having their children's photos miscategorized as gorillas is harmful? We know what biases people say offend them already, there's no evidence fixing them is harmful, but a non-zero risk that not fixing them is harmful. I feel like what I'm encountering is a conservative bias against changing the status quo, "social engineering", and the like. It seems people don't like deliberate, non-organic, changes to the status quo (well, they don't tend to like organic ones either, like succeeding generations becoming say, more sexually liberal) Machine learning can create filter bubbles, echo chambers, feedback loops, and people may attribute more weight to answers provided by machines than people, and so having machine learning reinforce even more strongly, current cultural biases that we're already seeking politically to ameliorate seems prudent and pragmatic to try and balance.
belorn 8y ago
Even for the low hanging fruit there is some concerns that the bias correction will only be made based on view of those currently in power at Google. They might want to make sure that translation AI will make careers like scientist, engineer, doctor to not get associated with male gender, but will Google do the same for crime and criminals? Google may want to steer YT recommendation AI away from the damsel in distress trope, but will it do the same for the bumbling dad trope? In current polarized climate there isn't much trust left that bias correction will be itself unbiased, or if it will reduce existing discrimination rather than doing the opposite.
repolfx 8y ago
It doesn't have anything to do with censoring your favorite conservative search result. But that's not correct. That's exactly what it's to do with. A big part of the philosophy of conservatism is to accept the world as you find it. Conservatives, at least in theory (not saying the Republicans are a great implementation of the philosophy), eschew large social engineering schemes, they eschew attempts to remould attitudes via manipulation of language and so on. These are all traits associated with the opposing end of the political spectrum. Think how important re-engineering people's thinking via language was in Orwell's 1984, for example. So now we have Google and related AI researchers announcing that when an ML model learns the actual fact that most doctors are male, this is "bias" and it needs to be "corrected". This is Orwellian. It's not at all biased, it's accurate. But because of some vague, handwavy assumption that if AIs believe gender ratios are 1:1 in medicine then somehow ... via some obscure chain of action ... more girls will become doctors, the basic reality of the matter is justified in being inverted. Or possibly, more boys will choose NOT to be doctors. Quite why this outcome is better than the existing status quo is never explained with any rigour - it's ideology. This is the very style of social engineering that conservatism rejects, on the grounds that it so often goes wrong or has unintended side effects. So whilst I am interested to see that Google is deciding to walk the walk here when it comes to AI and weapons, I nonetheless find their statement of moral principles to be quite chilling and frankly, it renders their most important products useless for me. Their take on this isn't news, but it's sad to see them doubling down on it. I do not wish to be subtly manipulated via AI driven translation or re-rankings into not believing things about the world that are true but upset radical feminists. It's especially sad for me because I worked there for so many years.
- cromwellian 8y ago
  Obtaining the correct categorization of something is not social engineering, it's returning a FACTUAL answer. I didn't ask the AI "are men more likely to be doctors" I asked "what is a doctor" Do you understand the difference? I don't expect it to tell me "Doctors are male individuals who...." No where in my original post did I say the AI has to lie about gender ratios, I would expect any AI to tell me also, the socio-historical-cultural perspective of the medical profession, and how it has changed over time. I am simply not interested in a neat little, static, steady state view of something that conveniently leaves out context. Do conservatives want to exist in a bubble where no criticism of the status quo is seen? If a machine learning algorithm learned a bad definition of conservative, that cast them as crypto-racists, you'd want it corrected wouldn't you? Even if a reading of conservative news sites comment forums gives you exactly that impression is likely true? > "Quite why this outcome is better than the existing status quo is never explained with any rigour - it's ideology." Right, progressives are the only ones with ideology, conservative positions are arrived at by cold, hard logic? Tell me why an African American slave boy, growing up in the 1800s, who learns the existing status quo that black people are property, and white people or not, is a worse outcome? Clearly, it's a better outcome for white people of the era, so any explanation for why it might be preferable to alter it has to argue that the status quo wasn't good for black people, or that it wasn't good for white people for some presumably economic reason. Maybe just maybe, the status quo isn't good for women? Maybe it would be good to ask them if it's ideology, or if changing cultural attitudes about what women are allowed to do, and capable of doing over the last 200 years has been a positive change in the status quo for them? BTW, Conservatives don't reject social engineering, they just reject social engineering they disapprove of. Social conservatives around the world, in concert with religion, have sought to engineer human behavior with appeal to cultural and religious piety, and in many cases, winning laws that enforce such behavior. We've had anti-sodomy laws, anti-booze laws, anti-miscegenation laws, all of them enacted by social conservatism. And what do you call religious proselytizing, if not social engineering? Trying to spread memes and infect and convert more people into a new way of thinking. Sometimes I feel that conservatives are against secular, humanist 'social engineering' the way Scientologists are against Psychotherapy, because it's competing in the same meme space. Society is a dynamic equilibrium. It is constantly evolving, sometimes it evolves purely organically, in a spontaneous order, and sometimes there are clusters and movements that boil over, and change arrives by deliberate persuasion. Google is a global, transnational company, that serves the entire world, 7+ billion people. It needs to reflect that diversity. And like I said, it absolutely cannot have AI that does stupid stuff like learn that black faces are gorillas.

minikites 8y ago

"Machine learning is like money laundering for bias" https://twitter.com/pinboard/status/744595961217835008 https://twitter.com/pinboard/status/744595961217835008

zawerf 8y ago

It's definitely a problem worth alleviating. But it is "triggering" because it is an open problem to determine whether a bias is harmful, even for human beings. So it becomes an impossible/unreasonable amount of extra work if you demand it as a prerequisite. For example, in your translation tool example, even a human translator would have trouble making the least offensive translation possible. She/he/(insert favorite pronoun here) would need to realize the audience is a young impressionable child who is about to base her entire world-view on whether there's statistically more of her gender in that one sentence of translation. For a machine learning algorithm to understand enough about human nature to not offend the parent of that child, you're better off waiting for AGI that can surpass human tact.

cromwellian 8y ago
You don't think a black person trying to use a photo management app and having their children's photos miscategorized as gorillas is harmful? We know what biases people say offend them already, there's no evidence fixing them is harmful, but a non-zero risk that not fixing them is harmful. I feel like what I'm encountering is a conservative bias against changing the status quo, "social engineering", and the like. It seems people don't like deliberate, non-organic, changes to the status quo (well, they don't tend to like organic ones either, like succeeding generations becoming say, more sexually liberal) Machine learning can create filter bubbles, echo chambers, feedback loops, and people may attribute more weight to answers provided by machines than people, and so having machine learning reinforce even more strongly, current cultural biases that we're already seeking politically to ameliorate seems prudent and pragmatic to try and balance.

belorn 8y ago

Even for the low hanging fruit there is some concerns that the bias correction will only be made based on view of those currently in power at Google. They might want to make sure that translation AI will make careers like scientist, engineer, doctor to not get associated with male gender, but will Google do the same for crime and criminals? Google may want to steer YT recommendation AI away from the damsel in distress trope, but will it do the same for the bumbling dad trope? In current polarized climate there isn't much trust left that bias correction will be itself unbiased, or if it will reduce existing discrimination rather than doing the opposite.

repolfx 8y ago

It doesn't have anything to do with censoring your favorite conservative search result. But that's not correct. That's exactly what it's to do with. A big part of the philosophy of conservatism is to accept the world as you find it. Conservatives, at least in theory (not saying the Republicans are a great implementation of the philosophy), eschew large social engineering schemes, they eschew attempts to remould attitudes via manipulation of language and so on. These are all traits associated with the opposing end of the political spectrum. Think how important re-engineering people's thinking via language was in Orwell's 1984, for example. So now we have Google and related AI researchers announcing that when an ML model learns the actual fact that most doctors are male, this is "bias" and it needs to be "corrected". This is Orwellian. It's not at all biased, it's accurate. But because of some vague, handwavy assumption that if AIs believe gender ratios are 1:1 in medicine then somehow ... via some obscure chain of action ... more girls will become doctors, the basic reality of the matter is justified in being inverted. Or possibly, more boys will choose NOT to be doctors. Quite why this outcome is better than the existing status quo is never explained with any rigour - it's ideology. This is the very style of social engineering that conservatism rejects, on the grounds that it so often goes wrong or has unintended side effects. So whilst I am interested to see that Google is deciding to walk the walk here when it comes to AI and weapons, I nonetheless find their statement of moral principles to be quite chilling and frankly, it renders their most important products useless for me. Their take on this isn't news, but it's sad to see them doubling down on it. I do not wish to be subtly manipulated via AI driven translation or re-rankings into not believing things about the world that are true but upset radical feminists. It's especially sad for me because I worked there for so many years.

cromwellian 8y ago
Obtaining the correct categorization of something is not social engineering, it's returning a FACTUAL answer. I didn't ask the AI "are men more likely to be doctors" I asked "what is a doctor" Do you understand the difference? I don't expect it to tell me "Doctors are male individuals who...." No where in my original post did I say the AI has to lie about gender ratios, I would expect any AI to tell me also, the socio-historical-cultural perspective of the medical profession, and how it has changed over time. I am simply not interested in a neat little, static, steady state view of something that conveniently leaves out context. Do conservatives want to exist in a bubble where no criticism of the status quo is seen? If a machine learning algorithm learned a bad definition of conservative, that cast them as crypto-racists, you'd want it corrected wouldn't you? Even if a reading of conservative news sites comment forums gives you exactly that impression is likely true? > "Quite why this outcome is better than the existing status quo is never explained with any rigour - it's ideology." Right, progressives are the only ones with ideology, conservative positions are arrived at by cold, hard logic? Tell me why an African American slave boy, growing up in the 1800s, who learns the existing status quo that black people are property, and white people or not, is a worse outcome? Clearly, it's a better outcome for white people of the era, so any explanation for why it might be preferable to alter it has to argue that the status quo wasn't good for black people, or that it wasn't good for white people for some presumably economic reason. Maybe just maybe, the status quo isn't good for women? Maybe it would be good to ask them if it's ideology, or if changing cultural attitudes about what women are allowed to do, and capable of doing over the last 200 years has been a positive change in the status quo for them? BTW, Conservatives don't reject social engineering, they just reject social engineering they disapprove of. Social conservatives around the world, in concert with religion, have sought to engineer human behavior with appeal to cultural and religious piety, and in many cases, winning laws that enforce such behavior. We've had anti-sodomy laws, anti-booze laws, anti-miscegenation laws, all of them enacted by social conservatism. And what do you call religious proselytizing, if not social engineering? Trying to spread memes and infect and convert more people into a new way of thinking. Sometimes I feel that conservatives are against secular, humanist 'social engineering' the way Scientologists are against Psychotherapy, because it's competing in the same meme space. Society is a dynamic equilibrium. It is constantly evolving, sometimes it evolves purely organically, in a spontaneous order, and sometimes there are clusters and movements that boil over, and change arrives by deliberate persuasion. Google is a global, transnational company, that serves the entire world, 7+ billion people. It needs to reflect that diversity. And like I said, it absolutely cannot have AI that does stupid stuff like learn that black faces are gorillas.