Slashdot Mirror
A New Bill Would Force Companies To Check Their Algorithms For Bias (theverge.com)
An anonymous reader quotes a report from The Verge: U.S. lawmakers have introduced a bill that would require large companies to audit machine learning-powered systems -- like facial recognition or ad targeting algorithms -- for bias. The Algorithmic Accountability Act is sponsored by Senators Cory Booker (D-NJ) and Ron Wyden (D-OR), with a House equivalent sponsored by Rep. Yvette Clarke (D-NY). If passed, it would ask the Federal Trade Commission to create rules for evaluating "highly sensitive" automated systems. Companies would have to assess whether the algorithms powering these tools are biased or discriminatory, as well as whether they pose a privacy or security risk to consumers.
The Algorithmic Accountability Act is aimed at major companies with access to large amounts of information. It would apply to companies that make over $50 million per year, hold information on at least 1 million people or devices, or primarily act as data brokers that buy and sell consumer data. These companies would have to evaluate a broad range of algorithms -- including anything that affects consumers' legal rights, attempts to predict and analyze their behavior, involves large amounts of sensitive data, or "systematically monitors a large, publicly accessible physical place." That would theoretically cover a huge swath of the tech economy, and if a report turns up major risks of discrimination, privacy problems, or other issues, the company is supposed to address them within a timely manner.
The Algorithmic Accountability Act is aimed at major companies with access to large amounts of information. It would apply to companies that make over $50 million per year, hold information on at least 1 million people or devices, or primarily act as data brokers that buy and sell consumer data. These companies would have to evaluate a broad range of algorithms -- including anything that affects consumers' legal rights, attempts to predict and analyze their behavior, involves large amounts of sensitive data, or "systematically monitors a large, publicly accessible physical place." That would theoretically cover a huge swath of the tech economy, and if a report turns up major risks of discrimination, privacy problems, or other issues, the company is supposed to address them within a timely manner.
When the facts say that men are on average stronger and taller than women, are the facts wrong? Some things are fact. Other things are uncomfortable facts. When facts conflict with beliefs (especially politically), which do you think will win?
Bias is favoring one thing over another. Which is what you want certain algorithms to do. I want Youtube to find stuff I like. I want Google to find pages that are relevant to me.
Not sure how you are going to tease out the "good" bias from the "bad" bias, though. To extend your example, if 90% of the people in Hong Kong looking for a famous concert pianist are trying to find Lang Lang, who is hugely popular there, he's going to come up pretty fast when looking for concert pianists in general, which is what you want. It means the algorithm is being biased against Helene Grimaud, which is fine, because she isn't what most people are looking for in Hong Kong. That doesn't mean she doesn't come up at all, it just means she's ranked lower in the search results.
My Other Computer Is A Data General Nova III.
Let's get some hard data showing the bias that is present in censorship. The US is more conservative than liberal:
https://news.gallup.com/poll/2...
Without algorithmic bias online media would lean conservative for the simple reason that the US has more conservatives than liberals. Yet somehow online platforms (Reddit, Facebook etc.) tilt overwhelmingly liberal.
This can only be a result of bias that has been put into algorithms and sanctioned.
I am currently reading "The Sum of Small Things." In the first chapter, the idea of different racial groups having different demand levels is shown through the data.
There are valid reasons for that difference in demand. However, to pretend it isn't there is to try to live in denial.
In the case, in the book, the increased demand by Blacks for conspicuous consumption goods, ceteris paribus, is based on the belief that many Blacks find it necessary, but often not the result of conscious decision making, to carry visible markers of the middle class because it is not assumed. Now, we can reject this conclusion. However, to reject the discussion because we reject the data gets us no closer to truth. instead, it moves us away from truth.
Unfortunately in todays world, political correctness wins - and you'll be banned for stating the facts.
When facts conflict with beliefs (especially politically), which do you think will win?
Especially when we are talking politically the answer is clearly beliefs. How else do you explain Trump and Brexit?
If we're using neutral data as an input and the system comes to it's own conclusions...doesn't that say something about the data set? Shouldn't we try to understand why the algorithm came to that conclusion instead of immediately jumping to "check your privilege" ?
Mod me down with all of your hatred and your journey towards the dark side will be complete!
Which has nothing to do with bias. Bias, in this context, is unwarranted assumptions. Men are on average stronger and taller than women, but a system which, say, ranks potential firefighter applicants using their gender as a factor instead of looking at their performance in the actual job is biased.
const int one = 65536; (Silvermoon, Texture.cs)
SJW, n: "Someone I don't like, and by the way I'm a fuckwit" - AC
Interestingly enough in your example, even if you removed the actual gender from the data, you'd probably still have a 'biased' selection algorithm.
This came up in some other scandal where an algorithm *tried* not to be racist by excluding race and ended up still very biased in a law enforcement context. Note that the algorithm was seemingly bogus for other reasons so its not the best example, but even if it was worknig correctly it still probably would have been biased and the bias would have been undeserved. Notably they looked at arrest records of the parents as an indicator, and if a biased system caused their parents to be arrested, then the system would gladly extend that bias to a new generation.
Which all points to a key problem of playing 'whack-a-mole' with various endpoints where bias manifests when the bias problem is a bit more systemic. If a field is unfairly excluding minorities or women, then you don't just wave a wand at the employers, you have to look at education and cultural upbringing and accept that correcting a balance problem may be a generational problem. Also make sure the people you think are being slighted actually want this kind of help, rather than elevating the state of thnigs they would rather do.
XML is like violence. If it doesn't solve the problem, use more.
Same for almost anything. Skin colour rarely matters, and given enough more direct data on factors that do matter, skin colour will have no predictive value, so the algorithm will ignore it.
If that were true then it wouldn't matter. However, either by nature or nurture, color matters. If it didn't Asians wouldn't be given penalties and other groups bonuses for college admissions. If we want to argue that race isn't important, and I don't think it is important, then we have to do away with the diversity quest and let it play out.
Algorithms have no intrinsic bias, they are just a huge set of algebraic equations or, depending how you mean it, the software implementation of said equations. Any bias you have in the results comes from the training dataset. Any "expert" who rants about how "algorithms have bias" is clueless.
Responsibility lays on whoever trains the system (a.k.a. optimizes for a certain space of data points), not on the math equations or the hardware implementing them.
Responsibility lays on whoever aims the gun and squeezes the trigger, not on the physics of launching a bullet from the gun, or on the gun that implements the said physics concepts.
Which has nothing to do with bias. Bias, in this context, is unwarranted assumptions. Men are on average stronger and taller than women, but a system which, say, ranks potential firefighter applicants using their gender as a factor instead of looking at their performance in the actual job is biased.
Sorry, but you haven't been listening to the left if you think the test for bias is about assumptions. Equal outcomes is very strongly being pushed as the measure for bias.
Do you have more males than females going into trades? That must mean a bias against females exists in the trades.
Do you have more Asians getting into STEM? That must mean a bias in favor of Asians in STEM.
Language is being redefined and weaponized to push people's agendas(I know, it always has). Today we have the definitions of equality, racism, bias, violence, assault and others being changed to better fit agendas. Racism being the grossest example because of it's importance and power. I grew up understanding racism to be discrimination based on race. Today though the push is on to redefine racism to be a combination of discrimination AND power. This turning the convenient trick then that 'whites' have all the power, so now only they can be racist, by definition.
the congress of the United States can not work on any law or regulation until the 12 appropriation bills that make up the budget of the United States of America are passed by congress and signed by the president.
;)
FYI the US government rarely does a budget anymore they are to busy doing useless political investigations and passing things that are just a waste of tax payers time and money.
One plus is the useless ness of government is bi partisan. The US government is now made up of DEMs and GOPers who think their government job is to do the bidding of their parties
Maybe administration/congress/president and their staffs should not get any pay checks either until they DO THEIR MAIN JOB!!!! a budget!!
Just my 2 cents
Even if it matters, it might not be fair. Let's say that a group of the population is on average more poor. Given that a person's information is somewhat noisy, a good machine learning algorithm that can determine that a person is a member of a poor group will give that person a bias towards poor. In other words, that person might have identical financial records as someone from another group and receive a different outcome for say a loan approval. This is a rational (improves accuracy) somewhat Bayesian decision by the algorithm, but most would say it is unfair.
Chris Mesterharm
Were you referring to the case of some software that was being used by judges to determine sentencing based on likelihood of recidivism? I do recall that particular case using data like parents' arrest record, along with a lot of other questions that had a higher likelihood of occurring in individuals from poorer communities, which has a strong racial correlation in many places.
Assuming the algorithm is appropriately designed, it should only matter if whether or not your parents being incarcerated is a good predictor of recidivism. If it isn't, the data would show that a bunch of black people who were arrested had children who didn't commit crimes, and that there were several black parents who were never arrested that did have children who committed crimes. I understand that someone could easily look at the data wrong and make terrible conclusions (see the gender pay gap as one common example) based on bad reasoning, but that's another matter if we're assuming that the algorithm was properly designed.
I think you'd have a stronger claim with the argument that arresting parents over trivial matters or non-violent crimes eroded the family structure in many African American communities which resulted in an increase in criminality. Studies that have explored this to that level of details support such reasoning. It isn't that black people commit more crime because they are black, it's that poor people from single-parent families commit more crime, and there happen to be a disproportionate number of black people that fall into that group. It's not the only factor, but we'd significantly reduce the problem by decriminalizing drugs.
Data science uses training data that often contains factors like race, sex, income, and education level that when included cause an algorithm to train to moderate or recommend people differently based on your assigned group.
But even if you leave out factors like race and sex, it is possible that the machine learning application figures it out for itself, and creates an internal pattern that happens to produce a good match to race or sex, or any other factor, and then discriminates based on that internal pattern.
If an investigator then looks at the result, it appears that the AI has certain biases.
I'll bet if you completely exclude specific criteria from being a factor for consideration because of some undesired bias that might occur around such information, the resulting decisions may still show bias.
File under 'M' for 'Manic ranting'
Same for almost anything. Skin colour rarely matters, and given enough more direct data on factors that do matter, skin colour will have no predictive value, so the algorithm will ignore it.
Exactly the opposite will happen. .let's pick example that won't brush POC people wrong as it disadvantages men, car insurance.
E.g
You let AI know the gender - it will figure men are more likely to get into crash.
If you stop feeding gender to AI, it will figure people named John are more likely to be causing trouble, than people named Julie.
You will have to play a long cat and mouse game cutting off information sources for the AI to a point of it becoming useless.
Just think about the whole "discrimination" issue.
Black and Latinos are poorer than White/Jewish/Asian americans.
Hence AI, created to optimize the strategy, figured it makes more sense to target the latter with ads.