Slashdot Mirror
Google Research Promotes Equality In Machine Learning, Doesn't Mention Age
An anonymous reader writes: New research from Google Brain examines the problem of 'prejudice by inference' in supervised learning -- the syndrome by which 'fairness through unawareness' can fail; for example, when the information that a loan applicant is female is not included in the data set, but gender can be inferred from other data factors which are included, such as whether the applicant is a single parent. Since 82% of single parents are female, there is a high probability that the applicant is female. The proposed framework shifts the cost of poor predictions to the decision-maker, who is responsible for investing in the accuracy of their prediction systems. Though Google Brain's proposals aim to reduce or eliminate inadvertent prejudice on the basis of race, religion or gender, it is interesting to note that it makes no mention of age prejudice -- currently a subject of some interest to Google.
If even machines come up with measurable differences between work performance of males and females, then I think giving them in average the same amount of money or the same promotions is discrimination. I'm all for giving a woman who performs just as well as a man the same money, but if there are additional risk factors like a pregnancy or when the parent has to raise child, the person usually prioritizes these things over work, so why should work not be allowed to prioritize that person over others who do not raise children or do not drop out for weeks and months out of some work-external reason.
I think that AIs, by definition, cannot have bias.
No. There is nothing in the "definition" of AI that prevents bias. AIs will be biased if the training data supports the bias. For instance, if the AI looks at loan default rates, it will conclude that blacks and Hispanics are worse credit risks than whites ... because they are. But discrimination in lending is still illegal even if it is supported by the facts, and even if it is determined indirectly by, say, zipcode, or given name.
You are highly overexagerating the level of "intelligence" of AI. The data going into a machine learning system is typically in the exact same format as what comes out. If you have a loan application application (sorry, couldn't resist myself) that predicts based on marital status and children, than the only type of data going in is long table with three columns; married (yes or no), children (yes or no) and repaid (yes or no). The AI is not going to get newspaper articles and infer all kinds of possibilities about what a marriage is. The only thing the AI knows about marital status is that status "yes" had different letters in it from status "no". The problem discussed here is that you cannot completely remove the data for "gender", as the combination of the data for "married" and "children" is not universally distributed amongst genders. Essentially, you cannot remove a bias unless all other data is completely independant of the data you want to remove.
Slashdot social media options: AIM, ICQ, Yahoo, Jabber and Mobile Text. Why no MySpace?
Yes, but being a single parent is a risk factor. You usually don't have as much time to focus on your job, etc. Or it can be the opposite: if you have a child, you want the best for them and maybe make extra sure you keep your current job, etc.
And about skin color, blacks have a larger unemployment rate than whites:
http://www.theatlantic.com/bus...
So you are not supposed to look at the employment status because due to this you might infer the skin color and apply racist bias? This is just totally nuts. Of course, you should not use skin color information to infer employment status, which would be racist, but using employment status information to make your loan decision should be possible, just as using information on whether you are a single parent or not.
You seem to work from the assumption that women and minorities are more likely to skip out of their bills.
You don't need to "assume" anything. You can just google the data.
Women are less likely to default on their mortgages.
Women are more likely to default on their student loans, partly because their degrees are more likely to be worthless so they earn less.
Blacks and Hispanics are more likely than whites to default on all types of loans.
Asians are less likely than whites to default.
well if the data backs up the claims, its not sexist, or racist
have you seen my sig? there are many others like it but none that are the same
The only "solution" will be if every living thing has the same result, so just ignore all values and hardcode the one output.
So g00gle found out that different groups really are different in a number of relevant factors and their conclusion was that evil cisgendered bigots when seeing inferior relevant attributes are going to automatically figure out an applicant is a protected minority and in their mind are somehow going to skip over the relevant reasons to discriminate and solely discriminate against based on them being a minority and the effect will somehow be distinguishable and worse than if they had just stuck with discriminating with the relevant reasons they already have.
I have no problems if the scales are tipped, just so long as they are in my favor.
If you want to be fair, instead of "order by score, race", you should "order by score, random". Ordering by race is racism plain and simple. Why not sort by shoe size? The answer is simple: shoe size (for most jobs) does not apply when analyzing for job qualifications. Your job qualifications are (mostly) not dependent on the color of your skin (with exceptions such as actors).
To help those out with a lack of understanding - Racisim(2): racial prejudice or discrimination.
the machine learning algorithm infers a difference which is real, but uncomfortable for us socially.
Let's assume that we can prove that the detected difference was in the case NOT introduced by human-created input-data bias.
I'll give an example: I'm left handed so I think I'm allowed to talk about this.
What if the system learns that left handed people in North America die a little earlier than right handed people.
And specifically that they die with higher frequency in car accidents.
(I'm pretty sure both statements are true above. Reasons for it are not definite, but for the first one, can include that many tools and affordances in society are designed to be easy for right-handers, so left-handers may interact poorly with them sometimes and sometimes that bites, For the second, it may be because a left handed driver who dozes off or becomes distracted tends to pull the steering wheel a little to the left, into oncoming traffic, Right-handers tend to pull to the right, onto the on-balance safer shoulder of the road.)
So does that mean its ok to increase life insurance premiums and automobile insurance premiums for left-handed people?
What kind of statistically valid discrimination IS ok? Any?
Then what do we do, in this day and age?
Where are we going and why are we in a handbasket?
The problem Google is describing isn't limited to a subset of arbitrary tribal factors society deems to be off limits.
Entire reason for existence of these systems is making prejudiced decisions about individuals based on statistical evidence.
You can spend all day filtering out things that will get you sued or attract bad press but this doesn't address core fact these systems are intended to make prejudiced judgments about individuals based on statistical experience and evidence.
Being prejudiced can be practically helpful in some contexts but don't pretend that isn't what your doing, don't confuse it for fairness and don't bother making up a bunch of mystical bullshit about how your dataset or programmers are biased. Prejudice is the raison d'etre of these systems. It is what they are designed to do.
I am not sure I agree. If the data says that $minority group is more violent then $non-minority, it may be statically true for a given set of statistics but we all (should) know that correlation is not causation and it may be that $minority group on average lives in a more dangerous place. Higher insurance rates for $minority group members would be racist, but charging higher rates for people (with out regard to race) living in a dangerous place would not be racist.
Causation is irrelevant in terms of insurance. The only thing that matters is accurately modeling risk. An algorithm doesn't have to know the reasons why kids are more likely to smash up their parents cars. It is only relevant that kids smash up their parents cars.
How can there be "prejudice" if the system _does not have cognition_? It just approximates a function. If a woman is less (or more likely) to default on a loan, it'll just say so, SJWs be damned. That's why women see ads for shoes even if they never disclosed that they are women to Google. That's also why they see fewer ads for engineering positions (women are statistically much less likely to be interested in engineering fields).
It's a function approximation problem, and this happens to be the function that the real world data seems to support. Now you want to wreck it for some kind of affirmative action, thus decreasing its accuracy and driving an agenda of what you think the world should look like, rather than what it actually is.