4 ms·

That example (https://research.google.com/bigpicture/attacking-discrimination-in-ml/ https://research.google.com/bigpicture/attacking-discriminat...) isn't part

by _Nat_ 6y ago

That example (https://research.google.com/bigpicture/attacking-discrimination-in-ml/ https://research.google.com/bigpicture/attacking-discriminat...) isn't particularly compelling. As their example shows, there're two traditional approaches: maximize profit and group-unaware. * Maximizing profits gets the most profit, but treats people differently based on their group. * Group-unaware treats everyone the same regardless of their group, but can generate far less profit. The example presents two alternatives: "demographic parity" and "equal opportunity". Presumably the authors would argue that these may be superior choices because they generate nearly as much profit as profit-maximization while having a plausible argument for being socially responsible. This seems a bit off. Fundamentally, we might say that there're 2 kinds of discrimination: fair and unfair. For example, it might be fair to discriminate for objective reasons, but it'd be unfair to discriminate for non-objective reasons (e.g., bigotry). The advantage of profit-maximization is that it takes full advantage of fair-discrimination while fully avoiding unfair-discrimination; the drawback is that it does discriminate. The advantage of group-unaware is that it fully avoids all discrimination; the drawback is that it sacrifices fair-discrimination, causing it to yield the lowest profits. The two alternatives proposed in that example seem to get the best of both worlds because they're basically just cloning profit-maximization, but with slight concessions to plausible-sounding criteria for equality to dodge perceptions of unfair-discrimination. Here're the tricks: * In "demographic parity", everyone has the same odds regardless of group. This would appear to be the same thing as profit-maximization if the risk/rewards were the same, but since it ignores them, it ends up being basically "profit maximization, but ignoring different risk/rewards". * In "equal opportunity", both groups get the same true-positive rate. This again seems to sacrifice some of the risk-vs.-reward information, but with a slightly different skew. So for the privileged group (Orange) vs. the disadvantaged group (Blue): * Profit maximization: $32,400 from 50 vs. 61 * Group-unaware: $25,600 from 55 vs. 55 * Demographic parity: $30,800 from 52 vs. 60 * Equal opportunity: $30,400 from 53 vs. 59 To me, that looks like 3 ways to discriminate, all yielding roughly the same profit and thresholds -- using either of the 2 proposed alternatives gives up a little bit of the profit in exchange for a pleasant-sounding rationale. What I dislike about this is that it seems entirely superficial. The proposed alternatives engage in roughly the same level of fair-discrimination (and none of them engage in unfair-discrimination, which wasn't given in the example at all) to generate roughly the same level of profit. --- > The conversation on this subject by lay folks is rife with statistical ignorance, and Google has done good work communicating and clarifying the conversational starting points. This is a hard issue because it takes ethical trade-offs and forces you to specify mathematically exactly how you want to handle inequality, which is a subject that most people haven’t thought through rigorously, and would rather hand-wave away with virtuous sound bites. That's exactly what this looks like! In this case, the virtuous sound-bites are "Demographic Parity" and "Equal Opportunity". They both worked out to be mostly the same as simple profit-maximization, but if someone in a disadvantaged group protests that they're being discriminated against, they'd probably find it difficult to follow the math far enough to sustain their complaint.

theptip 6y ago
I appreciate the detailed object-level analysis of that case, but my point was more intended at the meta-level — in order to have a conversation about AI bias, we need language and examples to start from. Google is building these fairness metrics into parts of its cloud ML toolchain, and is investing in peer-reviewed research in this area. That’s more than most other companies, and should be applauded. At the object level, I think you perhaps oversimplify with “ Fundamentally, we might say that there're 2 kinds of discrimination: fair and unfair.” What we think is fair is the crux of the whole issue, and there is not agreement on that distinction. For example, if we price risk at the “true” risk of default, and this means that fewer black people get loans, is that fair? Some think it is, and some think it is not, even though I hope most would agree there is no racist intent in the decision itself. The economic context in which these decisions are made already contains the impact of past racism, which can be perpetuated or or restituted by decisions made now. I also disagree with your characterization of these types of parity as sound-bytes. My definition of a sound-byte is a conceptually thin phrase that sounds good, but does not contain much content. These are the opposite; they are specific jargon with a precise meaning that is explained in detail in the article. You could call them “Foo” and “Bar” and their usefulness in the conversation would be the same. (Sure, “equal opportunity” has been deployed as a sound-byte elsewhere without any rigorous definition, but I hope you’d agree that this article at least fleshes out one possible concrete definition, even if you don’t think it is the right one. I’m sure they are others - and now we can further the field by writing them down, citing this piece, and making an objective critique!)
- _Nat_ 6y ago
  > My definition of a sound-byte is a conceptually thin phrase that sounds good, but does not contain much content. Yeah, same. And to my eye, this whole thing's paper-thin; it tested my suspension of disbelief to not dismiss it as an obvious joke or bad attempt at a PR stunt, and I'm honestly still undecided. Part of the issue is that its conclusions are absurd. We can invent ethical constraints to arrive at max-profit by pulling things out of thin air; the procedure's analogous to p-hacking. For funsies: 1. Come up with a huge portfolio of ethical-sounding constraints. Include basically any ethical-sounding argument imaginable, even if it seems stupid or bad for business (the numbers can be hacked; we just need the pretty words). 2. Come up with model-transform strategies. For example, instead of calculating the threshold, calculate the log-threshold or the threshold-entropy. (Again, don't think logically -- we're hacking the math, so just come up with random things.) 3. Come up with model-hybridization strategies. For example, construct a meta-model that uses 50% of the threshold + 25% of the inverse-log threshold + 20% of the inverse-entropy threshold + a flat 5% (because, hey, why not?). 4. Whenever we want to maximize profit, just do so with simple profit-maximization. No "ethical" constraints. 5. Repeat Step (4) for millions (or more) different hybridized models, combining all sorts of different model-transforms and ethical-constraints at random. Optimize each for approach to simple-maximization. 6. Stop upon finding an "ethical" solution that effectively equals simple profit-maximization. 7. Generate a description of the "ethical" solution, and write a report on how your company's using that particular ethical strategy as part of its latest, on-going campaign to promote social equity. 8. Just do simple profit-maximization. Or the "ethical" solution; whatever, they're literally the same thing. But if anyone asks, be sure to tell them it was for ethics. While we're at being ethical, we could even shift some paradigms to envision new modes of engagement with socioeconomically disadvantaged stakeholders to promote a synergistic realignment between our commitment to ethical outreach and disrupting the zeitgeist with our social media blitz to rapidly accumulate surprisingly liquid social capital, indelibly scrawled across the human blockchain of love and mutual respect, validated in the proof-of-work of billions of sapient hearts yearning to emerge from the trials-and-tribulations of our modern era into the welcoming bosom of a new tomorrow. But who cares about pointlessly defining jargon for logical inconsistencies, though? It's hallow and meaningless; at most, it might fool people before they realize that it's logically inconsistent, but that'd just be lies. It doesn't seem to have an actual use.

by _Nat_ 6y ago

theptip 6y ago

I appreciate the detailed object-level analysis of that case, but my point was more intended at the meta-level — in order to have a conversation about AI bias, we need language and examples to start from. Google is building these fairness metrics into parts of its cloud ML toolchain, and is investing in peer-reviewed research in this area. That’s more than most other companies, and should be applauded. At the object level, I think you perhaps oversimplify with “ Fundamentally, we might say that there're 2 kinds of discrimination: fair and unfair.” What we think is fair is the crux of the whole issue, and there is not agreement on that distinction. For example, if we price risk at the “true” risk of default, and this means that fewer black people get loans, is that fair? Some think it is, and some think it is not, even though I hope most would agree there is no racist intent in the decision itself. The economic context in which these decisions are made already contains the impact of past racism, which can be perpetuated or or restituted by decisions made now. I also disagree with your characterization of these types of parity as sound-bytes. My definition of a sound-byte is a conceptually thin phrase that sounds good, but does not contain much content. These are the opposite; they are specific jargon with a precise meaning that is explained in detail in the article. You could call them “Foo” and “Bar” and their usefulness in the conversation would be the same. (Sure, “equal opportunity” has been deployed as a sound-byte elsewhere without any rigorous definition, but I hope you’d agree that this article at least fleshes out one possible concrete definition, even if you don’t think it is the right one. I’m sure they are others - and now we can further the field by writing them down, citing this piece, and making an objective critique!)

_Nat_ 6y ago
> My definition of a sound-byte is a conceptually thin phrase that sounds good, but does not contain much content. Yeah, same. And to my eye, this whole thing's paper-thin; it tested my suspension of disbelief to not dismiss it as an obvious joke or bad attempt at a PR stunt, and I'm honestly still undecided. Part of the issue is that its conclusions are absurd. We can invent ethical constraints to arrive at max-profit by pulling things out of thin air; the procedure's analogous to p-hacking. For funsies: 1. Come up with a huge portfolio of ethical-sounding constraints. Include basically any ethical-sounding argument imaginable, even if it seems stupid or bad for business (the numbers can be hacked; we just need the pretty words). 2. Come up with model-transform strategies. For example, instead of calculating the threshold, calculate the log-threshold or the threshold-entropy. (Again, don't think logically -- we're hacking the math, so just come up with random things.) 3. Come up with model-hybridization strategies. For example, construct a meta-model that uses 50% of the threshold + 25% of the inverse-log threshold + 20% of the inverse-entropy threshold + a flat 5% (because, hey, why not?). 4. Whenever we want to maximize profit, just do so with simple profit-maximization. No "ethical" constraints. 5. Repeat Step (4) for millions (or more) different hybridized models, combining all sorts of different model-transforms and ethical-constraints at random. Optimize each for approach to simple-maximization. 6. Stop upon finding an "ethical" solution that effectively equals simple profit-maximization. 7. Generate a description of the "ethical" solution, and write a report on how your company's using that particular ethical strategy as part of its latest, on-going campaign to promote social equity. 8. Just do simple profit-maximization. Or the "ethical" solution; whatever, they're literally the same thing. But if anyone asks, be sure to tell them it was for ethics. While we're at being ethical, we could even shift some paradigms to envision new modes of engagement with socioeconomically disadvantaged stakeholders to promote a synergistic realignment between our commitment to ethical outreach and disrupting the zeitgeist with our social media blitz to rapidly accumulate surprisingly liquid social capital, indelibly scrawled across the human blockchain of love and mutual respect, validated in the proof-of-work of billions of sapient hearts yearning to emerge from the trials-and-tribulations of our modern era into the welcoming bosom of a new tomorrow. But who cares about pointlessly defining jargon for logical inconsistencies, though? It's hallow and meaningless; at most, it might fool people before they realize that it's logically inconsistent, but that'd just be lies. It doesn't seem to have an actual use.