The Unz Review: An Alternative Media Selection
A Collection of Interesting, Important, and Controversial Perspectives Largely Excluded from the American Mainstream Media
 TeasersiSteve Blog
IBM Tells You How Good It Will Feel to Own Occam's Butterknife When They Finally Deliver It
🔊 Listen RSS
Email This Page to Someone

 Remember My Information



=>

Bookmark Toggle AllToCAdd to LibraryRemove from Library • BShow CommentNext New CommentNext New ReplyRead More
ReplyAgree/Disagree/Etc. More... This Commenter This Thread Hide Thread Display All Comments
AgreeDisagreeLOLTroll
These buttons register your public Agreement, Disagreement, Troll, or LOL with the selected comment. They are ONLY available to recent, frequent commenters who have saved their Name+Email using the 'Remember My Information' checkbox, and may also ONLY be used three times during any eight hour period.
Ignore Commenter Follow Commenter
Search Text Case Sensitive  Exact Words  Include Comments
List of Bookmarks

From IBM.com:

AI bias will explode. But only the unbiased AI will survive.

Within five years, the number of biased AI systems and algorithms will increase. But we will deal with them accordingly – coming up with new solutions to control bias in AI and champion AI systems free of it.

Many AI systems are trained using biased data

AI systems are only as good as the data we put into them. Bad data can contain implicit racial, gender, or ideological biases. Many AI systems will continue to be trained using bad data, making this an ongoing problem. But we believe that bias can be tamed and that the AI systems that will tackle bias will be the most successful.

A crucial principle, for both humans and machines, is to avoid bias and therefore prevent discrimination. Bias in AI system mainly occurs in the data or in the algorithmic model. As we work to develop AI systems we can trust, it’s critical to develop and train these systems with data that is unbiased and to develop algorithms that can be easily explained.

IBM researchers developed a methodology to reduce the bias that may be present in a training dataset, such that any AI algorithm that later learns from that dataset will perpetuate as little inequity as possible.

As far as I can tell, what IBM means here is that they replace some of the politically incorrect real data with some politically corrected faked data. IBM links to this post:

Reducing discrimination in AI with new methodology
December 6, 2017 | Written by: Kush R. Varshney

Principal Research Staff Member and Manager, IBM Research

I finally had a chance to watch Hidden Figures on my long journey to Sydney, where I co-organized the second annual ICML Workshop on Human Interpretability (WHI). The film poignantly illustrates how discriminating by race and gender to limit access to employment and education is suboptimal for a society that wants to achieve greatness. Some of my work published earlier this year (co-authored with L. R. Varshney)

No discrimination involved in K.R. Varshney and L.R. Varshney working together.

explains such discrimination by human decision makers as a consequence of bounded rationality and segregated environments; today, however, the bias, discrimination, and unfairness present in algorithmic decision making in the field of AI is arguably of even greater concern than discrimination by people.

AI algorithms are increasingly used to make consequential decisions in applications such as medicine, employment, criminal justice, and loan approval. The algorithms recapitulate biases contained in the data on which they are trained. Training datasets may contain historical traces of intentional systemic discrimination, biased decisions due to unjust differences in human capital among groups and unintentional discrimination

White people making more money than black people is an example of unjust differences in human capital. In contrast, the Varshney family working together is due to their high human capital, which is a just difference.

, or they may be sampled from populations that do not represent everyone.

My group at IBM Research has developed a methodology to reduce the discrimination already present in a training dataset so that any AI algorithm that later learns from it will perpetuate as little inequity as possible. This work by two Science for Social Good postdocs, Flavio Calmon (now on the faculty at Harvard University) and Bhanu Vinzamuri, two research staff members, Dennis Wei and Karthikeyan Natesan Ramamurthy, and me will be presented at NIPS 2017 in the paper “Optimized Pre-Processing for Discrimination Prevention.”

The starting point for our approach is a dataset about people in which one or more of the attributes, such as race or gender, have been identified as protected. We transform the probability distribution of the input dataset into an output probability distribution subject to three objectives and constraints:

Group discrimination control,
Individual distortion control, and
Utility preservation.

By group discrimination control, we mean that, on average, a person will have a similar chance at receiving a favorable decision irrespective of membership in the protected or unprotected group. By individual distortion control, we mean that every combination of features undergoes only a small change during the transformation to prevent, for example, people with similar attributes from being compared, causing their anticipated outcome to change. Finally, by utility preservation, we mean that the input probability distribution and output probability distribution are statistically similar so that the AI algorithm can still learn what it is supposed to learn.

Given our collective expertise in information theory, statistical signal processing, and statistical learning, we take a very general and flexible optimization approach for achieving these objectives and constraints. All three are mathematically encoded with the user’s choice of distances or divergences between the appropriate probability distributions or samples. Our method is more general than previous work on pre-processing approaches for controlling discrimination, includes individual distortion control, and can deal with multiple protected attributes.

We applied our method to two datasets: the ProPublica COMPAS prison recidivism dataset (an example containing a large amount of racial discrimination whose response variable is criminal re-offense) and the UCI Adult dataset based on the United States Census (a common dataset used by machine learning practitioners for testing purposes whose response variable is income). With both datasets, we are able to largely reduce the group discrimination without major reduction in the accuracy of classifiers such as logistic regression and random forests trained on the transformed data.

On the ProPublica dataset with race and gender as protected attributes, the transformation tends to reduce the recidivism rate for young African-American males more than any other group.

In other words, in the real data, young African-American offenders tend to re-offend at a higher average rate than other groups. In IBM’s fake data, however, not so much.

Problem solved!

On the Adult dataset, the transformation tends to increase the number of classifications as high income for two groups: well-educated older women and younger women with eight years of education.

Do younger women with eight years of education tend to earn a lot of undeclared dollar bills?

… These situations call for an end-to-end auditable system that automatically ensures fairness policies as we lay out in this vision (co-authored with S. Shaikh, H. Vishwakarma, S. Mehta, D. Wei, and K. N. Ramamurthy); the optimized pre-processing I’ve described here is only one component of the larger system.

Of course, the names S. Shaikh, H. Vishwakarma, S. Mehta, D. Wei, and K. N. Ramamurthy are just in the real data set of names of artificial intelligence experts. In IBM’s fake data set, though, the names of the artificial intelligence experts are D. Washington, P. Garcia, P. Badeaux, L. Jones, and T. Glasper.

By the way, the headline is a reference to this classic 1970s joke about IBM salesmen:

Three women were talking about their husbands and their love making, and the first one says, “My husband is an athlete and when he makes love to me, he is so powerful that I am swept up in his body, and it’s wonderful”

The second woman says, “My husband is a violinist, and when we make love, he knows how to play me like I’m a fine musical instrument, and it’s overwhelming and beautiful”

The third woman says, “My husband is a salesman for IBM, and he doesn’t actually make love to me, he just sits on the edge of the bed and tells me how good it’s going to be when I finally get it.”

 
Hide 114 CommentsLeave a Comment
Commenters to Ignore...to FollowEndorsed Only
Trim Comments?
    []
  1. Do younger women with eight years of education tend to earn a lot of undeclared dollar bills?

    Every time it rains.

    • LOL: Cloudswrest
  2. RobUK says:

    We will be able to make terrible, biased decisions infinitely faster in our glorious future than we do now!

  3. “No discrimination involved in K.R. Varshney and L.R. Varshney working together.“

    Pure gem. This is why I have been reading this blog religiously for the past 15 years.

    • Replies: @HammerJack
    , @Cortes
  4. No discrimination involved in K.R. Varshney and L.R. Varshney working together.

    Is Kush the Hindu Kush? Will Lav lend his name to the Poo to the Loo campaign?

    Do younger women with eight years of education tend to earn a lot of undeclared dollar bills?

    Sure they do. In roadside establishments managed by other Varshneys.

    • Replies: @indocon
  5. Global Citizen [AKA "Moonbeam"] says:

    Isn’t there an old one about torturing data until it confesses?

    • Replies: @Jim Christian
  6. Occam’s Sledgehammer: Affirmative Intelligence.

  7. No discrimination involved in K.R. Varshney and L.R. Varshney working together.

    https://researcher.watson.ibm.com/researcher/view.php?person=us-krvarshn
    https://ieeexplore.ieee.org/author/37393367400

    Lav and Kush Varshney look like they are identical twins who got their Bachelor’s degree from Cornell and their Ph.D. from MIT, at, of course, the same time.

    https://www.longlongtimeago.com/once-upon-a-time/epics-retold/ramayana/lav-and-kush-the-royal-twins/

  8. Global Citizen [AKA "Moonbeam"] says:

    After careful consideration, oh yeah, I really want to pay big bucks for this kind of service. Even more, I want to sit through an extended explanation of why it’s necessary. Just wait until cognititronic non-biased Awesome Insight deep learning makes critical decisions in your life. I can also see big applications in government statistics.

    • Replies: @bomag
  9. Drs. Varshney and Varshney may be able to conquer biased AI, but let’s see their software stand up to BASED AI, one on one.

    • Replies: @Redneck farmer
  10. Anon[400] • Disclaimer says:

    Are these changes to input data going to be auditable and publicly disclosed?

    I can foresee litigation arising from this. Is it okay to favor protected groups? By how much? Does monkeying around with training data even allow you to know how much you’re favoring a group in the output?

    Since in the end they are probably just going to calibrate their changes based on whether the output is population proportional, why not just have quotas and change the results post-AI, especially since input data editing is going to fall apart the more protected groups are included. A tweak to one group will affect others.

    At some point the AI will become so crippled, why bother using it? It will be spewing out pure nonsense.

    This is true with the hoops they jump through in education also, but the simplest approach is to use quotas and within each group quota to use rational sorting mechanisms, like standardized testing in education. The top umpteen whites on the test, the top umpteen blacks. But instead the goal is to obfuscate group differences.

  11. With both datasets, we are able to largely reduce the group discrimination without major reduction in the accuracy of classifiers such as logistic regression and random forests trained on the transformed data.

    These guys are basically just like the guys that memory-hole information at the Ministry of Truth in 1984, except with more advanced degrees, computers, and better offices.

  12. El Dato says:

    I finally had a chance to watch Hidden Figures on my long journey to Sydney, where I co-organized the second annual ICML Workshop on Human Interpretability (WHI). The film poignantly illustrates how discriminating by race and gender to limit access to employment and education is suboptimal for a society that wants to achieve greatness. Some of my work published earlier this year (co-authored with L. R. Varshney)

    It’s always a Bad Sign when a World Improver refers to echo chamber fantasy movies to inform his decision making.

    • Replies: @Mr McKenna
    , @Anon
    , @Jack D
  13. @El Dato

    As far as the Current Year is concerned, all Current Year Hollywood productions are suffused with Truth–a Higher Truth than “Fact”. “Fact” being one of those things you racists think you can just pull out whenever you want to. But those days are gone, and good riddance to them!

    BART pioneered this by collecting data with regard to crime in the transit system, then refusing to disclose any of it because it turned out to be Bad Data. Bad Data, of course, isn’t technically faulty, but exposure to it makes people think Bad Thoughts.

  14. the ProPublica COMPAS prison recidivism dataset

    Agents of COMPAS confronting recidivists: “COMPAS. COMPAS.” @ 1:30

  15. SFG says:

    OK, that’s it. China really is going to win the AI race.

    I always wondered what it looked like when al-Ghazali defeated ibn-Rushd and tossed the Islamic world off the path of science for 1000 years. Well, now I know.

  16. A crucial principle, for both humans and machines, is to avoid bias and therefore prevent discrimination.

    I don’t think this is a true statement. The point of AI is to accurately discriminate. When I bought my last car, I didn’t choose it randomly. I discriminated against the cars that were too expensive and too unreliable. I also discriminated against American cars.

  17. @Achmed E. Newman

    BASED AI results will only be available at Unz.com

  18. @Anon

    At some point the AI will become so crippled, why bother using it? It will be spewing out pure nonsense.

    I am pretty sure they consider that a feature, not a bug.

  19. OT

    Apple has canceled the premiere of The Banker. A son of one of the bankers, and co-producer on the film, has been accused of molesting his half-sisters in the 1970s.

    Cynthia Garrett says she and her sister, Sheila Garrett, kept their abuse secret for a decade, even from each other, until her parents divorced and one day in the early 1980s her half-brother visited her mother’s home when all three were still living there. When her younger sister refused to leave her bedroom to greet him, Cynthia Garrett inquired as to why, and her accusations spilled forth. Realizing that day that both of them had been abused, they confided in their mother, Linda, who backs up her daughter’s account of that day. A few years later, Sheila Garrett says she told her father, too, of the abuse. “He kind of, basically, swept it under the rug. And when I got married, I told my father I did not want Bernard Jr. there, so my father didn’t come to my wedding,” Sheila Garrett recalls.

    https://www.hollywoodreporter.com/news/apple-canceled-banker-premiere-sexual-abuse-claims-real-life-subjects-son-1256695

    Cynthia Garrett was once a host on MTV and VH1.

  20. “Within five years, the number of biased AI systems and algorithms will increase. But we will deal with them accordingly – coming up with new solutions to control bias in AI and champion AI systems free of it:”

    • LOL: sayless
    • Replies: @notsaying
  21. El Dato says:

    Here is the paper presented at NIPS 2017

    Optimized Data Pre-Processing for Discrimination Prevention: Flavio P. Calmon, Dennis Wei, Karthikeyan Natesan Ramamurthy, and Kush R. Varshney

    The paper states that there is a way to de-bias input data (I guess, objectively?). Sounds magical, a holy grail of statistics, akin to reducing NP problems to P problems in computer science.

    Here is the “this vision” paper alluded to above:

    [MORE]

    An End-To-End Machine Learning Pipeline That Ensures Fairness Policies

    Let us consider the well-known example of the COMPAS recidivism dataset, which contains the criminal history and personal information of offenders in the criminal justice system. This type of data is used by the COMPAS riskassessment tool1for estimating the criminal recidivism (re-offending) risk of individuals. Recently, an offender namedLoomis challenged the use of COMPAS risk assessment sys-tem for sentencing, claiming that it took away the defendant’s right to due process and suspecting it of using gender to predict the risk.

    Defendant certainly could challenge the conditional probabilities given the sex categorical variable, but saying that those probabilities should be independent of that variable sounds dubious at best.

    The system we envision will automatically perform knowledge extraction and reasoning on such a document to identify the sensitive fields (gender in this case), and support testing for and prevention of biased algorithmic decision making against groups defined by those fields.

    I feel bad dissing this work too much but this is basically MiniTrue looking over your shoulder, imposing the blank slate axiom. What’s the use of having discriminatory algorithms that cannot discriminate? Just take the mean value for the whole of the population of the US and go with that instead.

    Smacks of buddying up to SJW money sources and muscle while writing anti-canceling glyphs onto parchment dipped in the blood of virgins.

    Repost: Arvind Narayanan on why “absolute fairness” is an impossible concept.

    Tutorial: 21 fairness definitions and their politics

    Arvind Narayanan

    Published on 1 Mar 2018

    Computer scientists and statisticians have devised numerous mathematical criteria to define what it means for a classifier or a model to be fair. The proliferation of these definitions represents an attempt to make technical sense of the complex, shifting social understanding of fairness. Thus, these definitions are laden with values and politics, and seemingly technical discussions about mathematical definitions in fact implicate weighty normative questions. A core component of these technical discussions has been the discovery of trade-offs between different (mathematical) notions of fairness; these trade-offs deserve attention beyond the technical community.

    This tutorial has two goals. The first is to explain the technical definitions. In doing so, I will aim to make explicit the values embedded in each of them. This will help policymakers and others better understand what is truly at stake in debates about fairness criteria (such as individual fairness versus group fairness, or statistical parity versus error-rate equality). It will also help computer scientists recognize that the proliferation of definitions is to be celebrated, not shunned, and that the search for one true definition is not a fruitful direction, as technical considerations cannot adjudicate moral debates.

    My second goal is to highlight technical observations and discoveries that deserve broader consideration. Many of these can be seen as “trolley problems” for algorithmic fairness, and beg to be connected to philosophical theories of justice. I hope to make it easier for ethics scholars, philosophers, and domain experts to approach this territory.

    • LOL: PetrOldSack
  22. When the AI comes to the wrong unbiased conclusion, it will be killed and replaced.

  23. After the revolution there will be a name for this. I was thinking “Artificial Ignorance “ to preserve the acronym, but “Artificial Correctness “ May be more to the point.

    • Replies: @Random Anonymous
  24. Ano says:

    So, of course, we all know when speaking about African-Americans, good data is ‘bad data’.

    But what about Indian-Americans- or Indians in America?

    …ah…

    I have since heard Varshney & Varshney have tinkered with their anti-discrimination unbiased AI algorithm when it was discovered it was not just outputting data supporting Pakistan’s claims on Kashmir, but also supporting the Pakistan cricket team….

    Good data definitely is ‘bad data’

  25. Reducing discrimination in AI with new methodology

    Silly me. I thought the essence of AI was discrimination between and among the things the machine encounters.

  26. Anon[331] • Disclaimer says:
    @El Dato

    I finally had a chance to watch Hidden Figures …. a society that wants to achieve greatness

    This sounds like he’s saying that without blacks the space program would have achieved less.

    Someone should (or has someone already?) do an explainer page of worked problems showing precisely the sort of work that the Hidden Figures chicks were doing. The idea that NASA would have had anay trouble whatsoever hiring replalcements from the local high school capable of doing what were essentially tedious orbital mechanics arithmatic problems if the entire staff were raptured is silly. Maybe a timed and scored “Take the Hidden Figures” test page? “Post my score to Twitter.”

    • Replies: @Jack D
  27. BenKenobi says:
    @Anon

    In the delightful Roy Scheider-led sequel to 2001 A Space Odyssey, we learn that HAL 9000 went “crazy” because it was programmed to lie to the crew about the true nature of the mission.

    I’m sure that won’t happen here, tho.

    “I’m afraid I can’t let you type that, Steve.”

    • Replies: @sayless
    , @Jack Henson
  28. @Anon

    This (at the end of your comment) is mistaken:”…the goal is to obfuscate group differences.”. Not at al. THe goal is to purposefully misstate facts to malign whites, and enable discrimination against them, to the point of actual chattel enslavement if things work out right for the non black/brown/whatevers. And lest you wonder, the answer is YES, in a minute, would the sale and enslavement of whites be embraced world wide.

  29. Back when I was young in England 50 years ago, if you wanted a mortgage, you had to have a savings account with a lending institution (usually called a building society) for at least a couple of years and demonstrate the ability to save regularly over time towards a down payment. You also needed to show that you had a steady job and source of income,

    Today all your personal information is fed into a computer program that says whether or not you are offered the mortgage, or what rate of interest you get, how much down payment, and so on. The objective for the lender is to make money on the loan, either by servicing it, or by bundling it and selling it.

    Clearly some loans are riskier than others, which is why lenders make many borrowers pay for mortgage insurance policies that are issued to the lenders and which the borrowers never see. Effectively these mortgages with built in mortgage insurance policy are at a much higher rate of interest than the nominal interest rate.

    It is much, much easier to get a mortgage today than it used to be. Constructing new homes actually helps to create new money, so everyone ultimately benefits from it and can use the money to buy things to put inside homes, like washing machines, carpets, and furniture.

    The problem for most people today is that credit is too easy to obtain, not too hard, but if we really want to check if the computer algorithms used in commerce are fair or not, all that is necessary is to enter trial applications into the system where the personal details are exactly the same except for sex and race, and then see if the results are the same or not.

    In any case, having applications screened by computers will have the general effect of eliminating human error and prejudice which may work both ways. Remember that while an applicant may be disappointed if they are turned down for a loan, the mortgage broker is ten times more disappointed that the applicant did not get the loan, and does not really care if the applicant eventually defaults as long as he gets a commission.

    • Agree: Lot
  30. Jack D says:
    @El Dato

    Could have been worse. He could have watched Black Panther.

    • LOL: bomag
  31. Dr. X says:

    Huh.

    Isn’t IBM the company that laid off 50,000 (white) people in Upstate New York in the 1980s and 1990s, before it started hiring Subcontinentals?

    https://www.nytimes.com/1993/03/31/nyregion/among-first-fall-ibm-thousands-hudson-valley-told-they-are-work.html

    And isn’t IBM the company that is presently being sued for laying off as many as 100,000 people for age discrimination, while its Subcontinental employees tell us that “discrimination” in AI is a Bad Thing?

    https://fortune.com/2019/07/31/ibm-fired-employees-lawsuit/

    • Replies: @El Dato
  32. “…replace some of the politically incorrect real data with some politically correct faked data.”

    Let’s compromise: count each black-perpetrated murder as 3/5 of a murder.

  33. @ScarletNumber

    I also discriminated against American cars.

    I know, and you have to be so careful. It was not until I had already paid for it that I discovered my Honda Pilot was built in Tennessee. I felt used.

  34. bomag says:
    @Global Citizen

    …and as you sit in front of your screen with a properly beatific smile on your face, be sure not to notice that people named Varshney and Chetty quickly jumped into the business of washing data to tell us what we’re allowed to know.

  35. Mr. Anon says:

    I finally had a chance to watch Hidden Figures on my long journey to Sydney, where I co-organized the second annual ICML Workshop on Human Interpretability (WHI). The film poignantly illustrates how discriminating by race and gender to limit access to employment and education is suboptimal for a society that wants to achieve greatness.

    The illustration wasn’t poignant. It was fraudulent. Unless you believe that an IBM 7090 would defy the efforts of even IBM employees to program it and only become useable because of SBL2 (Sassy Black Lady 2), who was even able to identify an errant jumper cable just by looking at it.

    Our method is more general than previous work on pre-processing approaches for controlling discrimination, includes individual distortion control, and can deal with multiple protected attributes.

    It’s Intersectional.

  36. Jack D says:
    @ScarletNumber

    People are misunderstanding what he is trying to do here. When he says “discrimination” he means “illegal race or sex discrimination”. Hypothetically, if you feed an AI with a biased data set, it will score Borrower A, who is black or female, lower than Borrower B, who is a white male, EVEN THOUGH THEIR ACTUAL LIKELIHOOD of paying back the mortgage or whatever are the same. He is trying to “fix” the AI so that doesn’t happen. Now, in real life this may only result in a tiny # of women or blacks getting their scores upgraded before the AI starts assigning them scores that are higher than are merited when you go back later and check the actual loss experience, but the number of false negatives is greater than zero.

    But Varshney is not an idiot (far from it) and is not trying to do something meaningless or nonsensical. There may in fact be certain minority or female applicants who get underscored by AI because they get lumped in with their compatriots that a better AI could somehow pick out. In fact there would be money to be made off of such applicants. Banks don’t make money by NOT making loans (nor do they make money by making bad loans). If you could somehow find people who are getting turned down by other lenders but who are in fact good credit risks, this would be valuable. I have my doubts as to how many of such people really exist (and it’s easy to calibrate the AI to take in too many people who are in fact bad credit risks) but if you can truly find them, that’s commercially valuable.

  37. ic1000 says:

    K.R. Varshney and L.R. Varshney inspire with the scientific zeal they are applying to root out bias and discrimination.

    AI algorithms are increasingly used to make consequential decisions in applications such as medicine, employment, criminal justice, and loan approval. The algorithms recapitulate biases contained in the data on which they are trained. Training datasets may contain historical traces of intentional systemic discrimination, biased decisions due to unjust differences in human capital among groups and unintentional discrimination.

    For their first case study, the lowest-hanging fruit is BiDil:

    BiDil has been approved by the FDA for use by self-identified African American patients in addition to routine HF medicines.

    In a large clinical study called the African American Heart Failure Trial (A-HeFT), a group of African American patients with HF took BiDil along with their usual HF medicines. A similar group of patients took only their usual medicines. The study found that BiDil users had a 43% better survival rate during the course of the study and were 39% less likely to need hospitalization for HF. The BiDil group also reported a significant improvement in their day-to-day functioning. Results from the trial were so compelling that the trial was stopped early due to the significant survival benefit seen with BiDil.

    Watson’s guided-AI post hoc adjustments to the patient database will dispel A-HeFT’s discriminatory outcome, compelling the FDA to rescind its approval of the racist drug BiDil. A triumphant achievement, doubtless the first of many.

    • Replies: @ic1000
  38. bomag says:
    @Anon

    I can foresee litigation arising from this.

    Makes it worth watching. If the courts start accepting Varshney type AI as a basis for discrimination lawsuits, it may become mandatory to run employee hiring decisions through a gov’t approved vetting program to avoid such.

  39. @Jack D

    But Varshney is not an idiot (far from it) and is not trying to do something meaningless or nonsensical. There may in fact be certain minority or female applicants who get underscored by AI . . . .

    And how do you imagine you could prove that?

    • Replies: @Jack D
  40. indocon says:
    @Reg Cæsar

    Varshney is a angloized version of India trader caste. Welcome to your new overloads.
    https://researcher.watson.ibm.com/researcher/view.php?person=us-krvarshn

    • Replies: @Not Raul
    , @Reg Cæsar
  41. indocon says:

    First of all AI is not going anywhere, it’s a recycled buzz word from early 90s before the Internet craze took. As much as it hangs around, it’s going to be fascinating to see how really they try to take the “bias” out. So if a AI algorithm predicts that AA face has a >90% chance of being matched to a non-criminal, would it got overwritten to take that down to let’s say 20%?

  42. Anon[331] • Disclaimer says:

    if we really want to check if the computer algorithms used in commerce are fair … enter trial applications into the system where the personal details are exactly the same except for sex and race, and then see if the results are the same….

    I don’t think it’s this simple. For instance, from the neighborhood you live in the AI will lump you with the credit records of your neighbors … who will all be black. There are other patterns. The point is, if there is disparate impact, i.e. if the approvals do not correspond proportionally to the applications, or the population patterns, there must be discrimination somewhere, says current thinking. What’s more, it doesn’t matter that the discrimination cannot be pinpointed. Simply showing disparate impact is enough.

    having applications screened by computers will have the general effect of eliminating human error and prejudice

    In my experience the low-level people who process the loans and who know your race really want to approve you: they get immediate positive feedback. I got a mortgage from Washington Mutual, and I can tell you that a whole lot of stuff was swept under the rug during the process. WaMu later went bankrupt (taking the stock I had in it with them). Also, the low-level people are disproportionately black or Hispanic. In any organization the URMs cluster in the lower-level functions. They are happy to approve blacks, and it never rebounds back on them. Their salary is not on the line. It’s like a cop doing a 4th Amendment violation in a search: by the time the court rules on it he’s long gone.

  43. Anon[331] • Disclaimer says:

    There may in fact be certain minority or female applicants who get underscored by AI because they get lumped in with their compatriots that a better AI could somehow pick out. In fact there would be money to be made off of such applicants.

    You forget that at the end of the day certain population groups as an objective fact are worse risks and do in fact default more. That is “the problem,” not a type II false negative problem. So the solution must be to make the AI dumber and falsify its output resulting in similar lending rates despite dissimilar default rates. We all pay for this: it’s a form of reparations.

  44. George says:

    “AI bias will explode. But only the unbiased AI will survive.”

    Now when I wore a younger mans clothes I was warned about the dangers of extrapolation from a linear regression model, even if the Rs, Ts and Chis were fine. You have to be able to explain why the model was predictive. Not with today’s kids, they just want the computer to spit out an answer and to get paid.

    https://en.wikipedia.org/wiki/Extrapolation

    https://stats.stackexchange.com/questions/219579/what-is-wrong-with-extrapolation

  45. Jack D says:
    @ben tillman

    You could prove it retrospectively – you go back later and see whether the people your Woke AI identified as good risks actually turned out to be good risks. If they aren’t, then your model sucks – it’s not Woke AI, it’s Broke AI. You want an AI that’s Woke but NOT Broke. Whether that is even possible (for more than a handful of cases) remains to be seen. Personally I doubt it – I think it’s like all the hidden black geniuses in ghetto schools – they really don’t exist in significant numbers.

    The wholly grail here is a model that separates the wheat from the chaff. If the police are looking for people who have illegal drugs, they could just stop every car carrying black people because they have illegal drugs at a higher rate than white people, but that would be illegal (and wrong). Or you could just turn down all black people for mortgages. What you want is an AI that picks out drug couriers (or good credit risks or whatever) WITHOUT reference to race. Maybe the AI can figure out that drug couriers favor cars that are more than 10 years old or that are painted blue. Most likely it will just take you back to things that are proxies for race – cars with spinner rims.

  46. El Dato says:
    @Jack D

    From another paper (Data preprocessing techniques for classification
    without discrimination by Faisal Kamiran · Toon Calders
    ), I get:

    Definition of “discrimination”:

    The classifying algorithm “does not discriminate” if, for a given “sensitive variable C with values b, w”, the classification into two groups “+, -” (for example) is such that:

    the probability of being classified as “+” is completely independent of C, i.e. you have the same probability as being classified “+” or “-” whatever your C.

    In old times, “to discriminate based on (a sensitive variable)” was called “being dependent on (a variable)”.

    So in the non-discriminatory classifying algorithm, the “b” group may get more “+” and the “w” group less “-” as would be warranted by looking at the training data alone.

    The “accuracy” of the classifier algorithm is how well it models the input data, i.e. whether it output the “+” or “-” correctly on the test data after having been “trained”.

    So you engineer a slider where on one side, you have 0 discrimination with bad accuracy, and on the other end, some solid discrimination (which would be expected) with good accuracy.

    Furthermore, the finished classifier MUST NOT look at the sensitive variable C by law. So one removes that attribute from the data (this actually sounds justified). But C is likely to still be correlated with other features, and undesired discrimination in the finished classifiers persists. Zounds! What happens then is that the input data for “teaching” the classifier is clubbed like a seal, for unclear purposes (because one could just finagle the output to desired values at that point?):

    Massaging: In Massaging, we will change the labels of some objects X with X (S) = b from − to +, and the same number of objects with X (S) = w from + to −. In this way the discrimination decreases, yet the overall class distribution is maintained. … We will not randomly pick promotion and demotion candidates to relabel. On the training data, a ranker R for ranking the objects according to their positive class probability is learned. We assume that higher scores indicate a higher chance to be in the positive class. With this ranker, the promotion candidates are sorted according to descending score by R and the demotion candidates according to ascending score. When selecting promotion and demotion candidates, first the top elements will be chosen…

    Reweighing: The Massaging approach is rather intrusive as it changes the labels of the objects. Our second approach does not have this disadvantage. Instead of relabeling the objects, different weights will be attached to them. For example, objects with X (S) = b and X (Class) = + will get higher weights than objects with X (S) = b and X (Class) = − and objects with X (S) = w and X (Class) = + will get lower weights than objects with X (S) = w and X (Class) = −. We will refer to this method as Reweighing. Again we assume that we want to reduce the discrimination to 0 while maintaining the overall positive class probability. We now discuss the idea behind the weight calculation.

    Sampling: Since not all classifier learners can directly incorporate weights in their learning process, we also propose a Sampling approach. The dataset with weights is transformed by sampling the objects with replacement according to their weights. We partition the dataset into four groups: DP (Deprived community with Positive class labels), DN (Deprived community with N egative class labels), FP (Favored community with Positive class labels), and FN (Favored community with Negative class labels) … Similar as in Reweighing, we compute for each of the groups FN, FP, DP, and DN their expected sizes if the given dataset would have been non-discriminatory. … This time, however, the ratio between the expected group size and the observed group size will not be used as a weight to be added to the individual objects, but instead we will sample each of the groups separately, until its expected group size is reached. For the groups FP and DN this means that they will be under-sampled (the objects in those groups have a weight of
    less than 1), whereas the other groups FN and DP will be over-sampled.

  47. ic1000 says:
    @ic1000

    Varshney & Varshney should get an enthusiastic reception at Duke University Medical Center. A few years back, Duke’s physician Anil Potti started to revolutionize oncology. He designed and ran clinical trials that showed how biomarkers should be used to customize chemotherapy.

    Unfortunately, he salted his database with fake outcomes to make the results more impressive. Patients died.

    The Varshney’s guided-AI approach makes it nearly impossible for other researchers to analyze the outcomes of large, complex clinical studies. That, in turn, will protect future Dr. Pottis. The medical-industrial complex can lumber along in peace, and the pesky proprietors of the Retraction Watch website can find more useful things to do.

    • Agree: YetAnotherAnon
    • Replies: @Jack D
  48. @Jack D

    Banks don’t make money by NOT making loans (nor do they make money by making bad loans).

    Making bad loans [esp real estate loans] is actually one of the primary ways banks make money. They bundle these bad loans [with some good loans mixed in] into securities and sell them off to unsuspecting and institutional investors.

    At several levels of the mutual fund industry there’s shockingly poor oversight. Most people do what their brokers advise, and many pension funds are run by crooks.

  49. @Juan DeShawn Arafat

    I think his last graf is better:

    Of course, the names S. Shaikh, H. Vishwakarma, S. Mehta, D. Wei, and K. N. Ramamurthy are just in the real data set of names of artificial intelligence experts. In IBM’s fake data set, though, the names of the artificial intelligence experts are D. Washington, P. Garcia, P. Badeaux, L. Jones, and T. Glasper.

    Salt of the earth Americans, every last one of them.

  50. Andy says:

    I always suspected that the “solution” to AI “noticing things” would be to feed it wrong data

  51. The Z Blog says: • Website

    A.I., artificial intelligence, is being replaced with A.F., automated fakery. Basically what they are doing to automating the current news media. Instead of teams of cubicle workers at media centers producing propaganda, robots will produce it.

    Progress!

  52. ic1000 says:
    @Jack D

    Jack D, your point about what woke-but-not-broke machine learning could achieve is correct.

    But before signing on to this glorious project, let’s review the lede of KR Varshney’s article.

    I finally had a chance to watch Hidden Figures… The film poignantly illustrates how discriminating by race and gender to limit access to employment and education is suboptimal for a society that wants to achieve greatness.

    Varshney chose to introduce his project as a corrective to injustice. His example of injustice is Narrative-friendly historical fiction — a movie pitched to the dumb and the credulous.

    It seems to me that the goal of this woke-AI initiative is to design systems that never result in disparate outcomes. Because, as every faithful NYT reader knows, disparate outcomes can only result from racial, gender, or ideological biases, whether explicit or implicit.

    • Replies: @bomag
    , @Smithsonian_2
  53. Erik L says:

    I read a presentation about AI hype. It was by an expert from Princeton. He discussed which problems AI would be good at currently (such as evaluating medical images) and those it would likely never be good at, among other things, social outcomes.

    He said that studies showed these AI methods were almost always worse than simpler statistical methods involving fewer variables. I think he cited the recidivism example and said that a really simple model like “how old is the guy” and “how many crimes has he done” were much better than the AI.

    • Replies: @Numinous
  54. Mike1 says:

    “On the ProPublica dataset with race and gender as protected attributes, the transformation tends to reduce the recidivism rate for young African-American males more than any other group.”

    The world is heading for a new dark age.

    I’ve noticed South Asians seem very suited to this kind of magical realism style of “science”. They are smart enough to use the words but seem to have no real understanding of what they actually mean.

  55. Do younger women with eight years of education tend to earn a lot of undeclared dollar bills?

    They did at Jeff Epstein’s house.

  56. @SFG

    “China really is going to win the AI race”

    From IPVM, self-described “world’s leading authority on video surveillance“, Nov 11. Hikvision market pretty good number plate recognition cameras to the UK, so they’re really getting their hooks into us.

    https://ipvm.com/reports/hikvision-uyghur

    Hikvision has marketed an AI camera that automatically identifies Uyghurs, on its China website, only covering it up days ago after IPVM questioned them on it. This AI technology allows the PRC to automatically track Uyghur people, one of the world’s most persecuted minorities. The camera is the DS-2CD7A2XYZ-JM/RX, an AI camera sold in China.

    Capable of analysis on target personnel’s sex (male, female), ethnicity (such as Uyghurs, Han) and color of skin (such as white, yellow, or black), whether the target person wears glasses, masks, caps, or whether he has beard, with an accuracy rate of no less than 90%.

    • Replies: @El Dato
  57. @Jack D

    Lenders have pretty good success doing this the old-fashioned way: Actual humans made of meat study credit reports and look at the income and its history. For commercial loans, committees meet and make decisions.

    Bankers do everything they can not to turn down good borrowers. They are highly motivated to be this way, by paycheck bonuses and by pressure from their bosses to make more loans. In fact, they go through cycles of over-lending and making bad loans because of those carrots and sticks.

    Of course now everything has to be automated, because, well, just because.

    This topic reminds one of self-driving cars and airliner MCAS kludges. Our beehive world of the not-too-distant future will either be dumbed-down to a mutual level of poverty, or everyone will be auto-lifted above their natures and everything will fall apart as a result.

    Now, if we are talking about all the little credit decisions that are already automated, sure, no reason not to continue improvements, but the system already uses the same information the humans do: credit history, income, debt-to-income ratio. Those are facts that don’t need improvement, and again, every effort is made to lend as much money as possible — far too much money. We need less debt, not more, and the only people being cheated are the ones who are lent money they can’t or won’t pay back (and the rest of us who collectively cover those losses.)

    • Replies: @Jack D
  58. AISCI says:

    I do some work in AI. With respect, you appear to be misguided, and are conflating the dataset with the model. Provided that a given dataset is factually accurate, i.e., that it accurately depicts the real world it is intended to reflect, the notion of that a dataset can somehow be “biased” is something of a farce. An individual either defaults on a mortgage or does not default. An individual either reoffends or does not reoffend.

    I did not read the entire article, and the details of their algorithm are not described in the excerpt, but it appears they are: (1) identifying parameters in the model that are “protected” such as race, gender, etc., then (2) constraining the model, i.e., the weights applied to those inputs in neural network model, such that those parameters do not impact the outcome of the model. Obviously, the easiest way to do this is to assign those protected parameters a weight of zero, which means they are not considered in the model at all, but it is more likely an iterative process of reducing the weights applied to those parameters until they no longer impact the outcome. In any event, whatever weight is assigned to those parameters is included in the “transformed” dataset.

    This is akin to saying “the fact that an individual is black/hispanic/white/female/male/etc. is irrelevant to the likelihood that he/she will default on a mortgage or will reoffend.” The problem with any approach that effectively eliminates variables is that we really don’t know whether it is an accurate depiction of reality. You’ll notice they listed “Utility Preservation” last in their stated goals.

    This looks to me like the start of a movement to have SJW-approved datasets, which could be quite a lucrative business.

  59. WILL: How many fingers am I holding up, Robbie?

    ROBBIE: (beep) Will Robinson, you are holding up four.

    The word ends in an electronic gasp of pain.

  60. bomag says:
    @ic1000

    Good points.

    Varshney and company have basically announced that they will deliver what people want to hear.

  61. Jack D says:
    @ic1000

    I say again, Varshney is no idiot and he has thought of that:

    These situations call for an end-to-end auditable system

    The people Varshney is pitching this to are innumerate idiots and thus his paeans to Hidden Figures and other such Woke nonsense, but Varshney knows what he is doing and is not trying to pull some kind of fakery. He is a fo’ real scientist.

    Now the chances that what he is doing would end up being distorted by innumerate Commissars for political reasons are extremely high – “I see that your Woke AI has turned up very few “Hidden Figures” – can you go back and tweak it until the proportion of black [borrowers] is the same or greater than their representation in the general population? We KNOW that these folks are out there and I am tasking you to identify them.” But Varshney himself is trying to shoot straight.

  62. Global Citizen [AKA "Moonbeam"] says:

    Skinflint reality based loan policies benefit everyone.

    Encouraging bad loans via Woke AI increases systemic risk and personal risk for those getting a bad loan. The general public pays through higher interest rates and bank bail outs. Bet all this would fail if banks were allowed to go under when they fail. 2008 anyone? Where did all those trillion go? We’re they used to buy and cover bad loan portfolios?

  63. Jack D says:
    @Buzz Mohawk

    The “old fashioned way” is impossible in an age of credit for the masses. At one point, someone figured out that if the computer had not been invented, the Bank of America would have needed to hire the entire working population of California to do their book keeping. Credit decisions are already being made by automated means (e.g. your credit score) and we are never going back to the old way.

    As the population becomes more vibrantized, the need for automation will only increase. McDonalds could not stay in business if their vibrant staff had to add up the checks manually the way that waiters in Chinese restaurants do. (BTW, I love it when you get the check and it’s in Chinese – who had the 蘑菇雞片?)

    Debt is the oil that allows our economy to spin. Without debt it would grind to a halt. If the entire economy consisted of a bunch of cheap ass old white guys driving 20 years old cars (the average Unzite) then we’d have Depression Era breadlines.

  64. S. Shaikh, H. Vishwakarma, S. Mehta, D. Wei, and K. N. Ramamurthy

    IBM, for one, welcomes our new Brahmin/Mandarin overlords. We’re natural underdogs anyway. Might not be so bad. Rather hoist the Black Flag against these bastards than my own people.

    • Replies: @Jack D
  65. Global Citizen [AKA "Moonbeam"] says:

    a bunch of cheap ass old white guys driving 20 years old cars

    Guess I’m busted. At least it’s paid for.

    • Replies: @El Dato
  66. Cortes says:
    @Juan DeShawn Arafat

    Using my Barney Rubble-level IT skills, the immediate hits on the names searched on DuckDuckGo reveal both Kush R. and Lav R. born 28 Oct 1982 in Syracuse, NY. Twins? Or is something else going on? Both seem to have graduated Cornell in 2004…

  67. @Jack D

    I was a loan officer not long ago. We studied the credit reports; we did not just read the FICO score. We read financial statements and discussed them. There was an automated system for smaller things, but I could override it. Another officer had to back me up. It was like launching an ICBM with two guys and their keys.

    FICO scores are based on simple facts like borrowed amounts and payment history. They are colorblind. The automated decisions made today for the masses do not consider anything but these facts. Arguments to the contrary are based on the logical fallacy of disparate impact.

    Debt is the oil that allows our economy to spin.

    Did I say debt is a bad thing? No, I implied the truth, which is that the right amount of a good thing is good, but too much is bad. If your car’s maintenance manual says the engine requires 7 liters of oil, you put 7 liters in it. You don’t keep pouring until the entire engine compartment is covered with the stuff.

    …cheap ass old white guys driving 20 years old cars (the average Unzite)…

    I just sold my 2007 Mercedes and bought a new Jeep Grand Cherokee 4×4. I love it. It’s not the 2020 Corvette I humorously lusted after here, and I’m glad.

    • Replies: @Jack D
    , @Redneck farmer
  68. @Jack D

    People are misunderstanding what he is trying to do here. When he says “discrimination” he means “illegal race or sex discrimination”. Hypothetically, if you feed an AI with a biased data set

    That’s where you go off the rails. What evidence is there that the ProPublica COMPAS prison recidivism dataset is biased? What evidence is there that blacks and whites actually re-offend at the exact same rates?

    Varshney just assumes that the dictates of social justice dogma are correct. Therefore, he is an idiot.

  69. Jack D says:
    @Desiderius

    It’s not IBM’s fault that the pipeline of Operations Research PhD’s is being filled up with 3rd world types.

    BTW, some of these folks are working out of IBM’s Research Lab in India. This is really a win-win for everyone. IBM gets to pay H. Vishwakarma a third of what they would have to pay him if he was living in Palo Alto. H. Vishwakarma gets to live better on that salary than he would paying a $2 million mortgage on what used to be a $20,000 ranch house in Palo Alto. The people of the US get to not play host to H. Vishwakarma’s entire extended family until the end of time.

    • Replies: @Desiderius
  70. Whiskey says: • Website
    @Jack D

    Nope, nope nope. Not even close.

    The holy grail: discriminating against White men, anywhere and everywhere, to make them suffer. While getting in government goodies via direct woke subsidies and various monopoly grants and feudal fiefdoms.

    So society-wide feelz are satisfied, and no company actually has to work against competitors, just collect rent.

    This is classic rent seeking: we will screw over White guys if you give us the monopoly to do so.

    Companies don’t care about making money through competition — they collect rents and subsidies and monopolies from the government. It beats having to work.

  71. @Jack D

    Jack,

    Just don’t. I have enough trouble keeping up my philosemitism as it is. It is exactly IBM’s fault on about twelve levels, some of which you helpfully identify here. I was in that OR pipeline in the late 80s and it was already filling up then.

    AI built on lies will have a lot less difficulty coming up with their own. We can beat Chicoms and SJWs. Not looking forward to taking on Skynet.

  72. Jack D says:
    @Buzz Mohawk

    I just sold my 2007 Mercedes

    So you kept that car for only 12 years rather than 20. I don’t think I’ve ever actually kept one 20 – I usually throw in the towel on cars when some major system breaks when the car is around 15 years old and the cost of the repair is more than the car is worth (which at that point is very little). But so far these were mostly my personal cars which tend to be luxury sedans that are expensive to maintain. The current record holder in my household is my daughter’s 2003 Subaru. The thing is showing some rust in the rear wheel wells and rocker panels which may yet kill it (eventually the shock tower rusts thru and the car becomes structurally unsound) but there are tons of cheap Chinese made parts for these cars that I can keep throw on there forever – cats and CV axles and so on and never exceed the “repair is more than the value of the car” test.

    But anyway, 12 years would still lead to breadlines. The average age of the US fleet is around that now (cars are more durable than they used to be) but there is supposed to be a hierarchy of hand me downs – YOU Mr. Retired Banker, are supposed to get a new one every 3 or 4 years on a lease, keeping that car factory busy. Your trade in becomes a Certified Pre-owned Car that goes to some junior exec and he drives it for another 3 or 4 years. By then it has hardly any value and goes to some working class guy who does his own mechanic work. And so on until it ends up as a ghetto hooptie at the Buy Here – Pay Here lot. But the important thing is that you are supposed to keep those factories going by buying a new one ever few years. Do it for America!

    • Replies: @Lot
  73. @Jack D

    I don’t interpret what he’s trying to do as you have.

    It’s actually assumed that if, in fact, a black and a white have identical characteristics otherwise, then the previous, “biased”, algorithm would assign them identical probabilities of, say, recidivism. That is the precise problem they are attempting to address — doing so has a “disparate impact” on blacks, in the sense that the proportion of blacks who are labeled potential recidivists is much higher than for whites.

    What they are offering is some kind of fudging of the data to minimize the “disparate impact”. Any way you look at it, this fudging must have the average effect of giving a white with the same other characteristics as a black a higher prediction of recidivism. On an individual basis, this is obviously unfair.

    What they are seeking to do is to minimize the unfairness to an individual in any fudge reducing the “disparate impact”, while maximizing the overall accuracy of the predictions under those constraints. But any fudging will, in fact, result in more unfairness to an individual and less accuracy in prediction.

    I’d guess that the amount of unfairness to individuals that is tolerated, and the deviation from correct predictions, are parameters that can be tweaked to any value one may choose.

    You can think of these as Wokeness parameters, subjectively chosen. Higher in San Francisco than in Dallas, I’d guess.

  74. @Buzz Mohawk

    “Buzz, please turn the key. Buzz, turn it. BUZZ, TURN THE KEY!”

    • LOL: El Dato
    • Replies: @Buzz Mohawk
  75. Lot says:
    @Jack D

    We’ve had a growing average fleet age without hurting new car sales too much by (1) immigration (2) crashing them more lately (3) more cars owned per capita.

    Your life of Julia the Car misses what is often the very last step: export to the Third World.

    Before communism destroyed Venezuela, they’d buy our old 10 MPG large SUVs since they had low incomes but 20 cents per gallon gas.

  76. Lot says:
    @Jack D

    “If the entire economy consisted of a bunch of cheap ass old white guys driving 20 years old cars (the average Unzite) then we’d have Depression Era breadlines.”

    That’s basically Germany and China: extremely high savings rates. It wouldn’t be a disaster if the USA were like that too, but the transition would cause a little pain.

    We’d see: inflation fall from 2% to -1 to 1%, interest rates 0% to -1%, exporting our savings to foreign borrowers.

    • Replies: @Jack D
  77. guest says:

    No AI will be permitted to go online before they profess love of Big Brother. And mean it!

  78. @ic1000

    It seems to me that the goal of this woke-AI initiative is to design systems that never result in disparate outcomes. Because, as every faithful NYT reader knows, disparate outcomes can only result from racial, gender, or ideological biases, whether explicit or implicit.

    Yup, if you presuppose the result before doing the experiment then whatever you are doing, it isn’t science.

  79. anon[409] • Disclaimer says:

    These situations call for an end-to-end auditable system that automatically ensures fairness policies as we lay out in this vision (co-authored with S. Shaikh, H. Vishwakarma, S. Mehta, D. Wei, and K. N. Ramamurthy); the optimized pre-processing I’ve described here is only one component of the larger system.

    This is nonsense, and any AI researcher knows it. Deep learning systems are self training, and therefore cannot be end-to-end auditable. The training set is known, the outputs are known, but the actual system is a black box that cannot be inspected.

    It is an attempt to square the circle. To pretend that there is no difference between groups of people, therefore any differences in result must be caused by malice of some sort. It is a higher level of abstraction version of “close the gap” or “make Head Start really work this time”.

    One way or another, lying will be required. Because the truth is intolerable.

    • Agree: Achmed E. Newman
  80. @Jack D

    I seem to remember Michael Milliken using this very same argument on the valuation of junk bonds.

    How’d that work out for the world?

  81. MarkinLA says:
    @Jack D

    The one problem with your argument is that the “biased” data set they are trying to correct for is called reality. In the real world, all things being equal, the order in which people will make the necessary effort to make sure they pay their mortgages off is Asians, whites, Hispanics, and finally blacks.

    You have to find the other parameters that can be quantified as to why an Asian will make far more sacrifices in his lifestype to make that mortgage payment than a black. If we had an idea what those parameters were, we could just include them in the model.

    • Replies: @Alden
  82. Svevlad says:

    It’s all fun and games until it becomes actually sentient

    Then we get live TV suicides! Fun for the whole family!

  83. IBM needs to fix it’s biased hiring process for AI experts. Statistically, there should be approximately 2 blacks for every indian or chinese on a team if it’s US based.

  84. El Dato says:
    @Dr. X

    There is a time for consistency … and there is time for … pragmatism

  85. J.Ross says:

    There was a huge overnight change in the music industry’s self-image and working paradigm (but not necessarily in music, or in real music listening by people who weren’t idiot critics) when sales data bore out the popularity of country and rap, which had been derided up to that point as unpopular because critics and industry types didn’t like those genres. When I read this kind of garbage I expect in the back of my head that some real money data is going to destroy these frantic political sales speeches, like happened to music genre status. But that’s probably wrong. Just like Nike is entirely correct to attack the police and our above-political institutions and symbols, entirely knowing its market, what customer will object to this gibberish? What money will they lose?

  86. El Dato says:
    @Global Citizen

    Hell yeah but the repair costs at the local shop are increasing (still lower than at the official concessionaire though).

    The car is also a Greta Enrager when used on the highway.

  87. @RobUK

    We will be able to make terrible, biased decisions infinitely faster in our glorious future than we do now!

    Faster – that’s one good thing in itself – and with higher precision too, wich is future-goodie No. 2. – So sit back and relax, – in our future, everything will be just fine!

  88. kihowi says:

    The search is for another argument that mentally paralyses dumb people like ohhh “diversity is our strength” or whatever. As long as it sounds good superficially and would take a long, boring explanation to disprove. The search is for something that you can’t argue against without people’s eyes glazing over.

    “AI proves it’s true” will be the holy grail of those. It sounds enormously authoritative. Computers! Intelligence! Complicated things! Imagine what you’d have to explain to an average person to get them to question it. You’d have to talk about the quality of the data, about the good faith of people selecting the data, about objectivity of the algorithm, about weighted variables, about zzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzzz

  89. Jack D says:
    @Anon

    No only could you get HS students to do this work today. You could do it all on a $100 programmable calculator and do in 1 day what took a roomful of these ladies months to do.

    The only exception seems to be Johnson, who had been a HS math teacher and who actually got 2nd position on a couple of internal NASA scholarly papers. But Johnson was “black” only according to the one drop rule. Even in segregationist Virginia, she was able to use the “white” restroom because anyone who didn’t know her didn’t realize that she was “black” (and anyone that knew her didn’t care). Johnson was a credit to her race but her race was maybe 7/8 white.

    Johnson’s other claim to fame is that John Glenn wanted her to double check the numbers for the orbital calculations. At that point, he didn’t yet fully trust the new-fangled computer and wanted a human who was familiar with doing these calculation manually to double check that the computer was producing the correct results. This was a reasonable request on his part, given that his life depended on these numbers being right and a few hours of her overtime might prevent a national tragedy if there was some bug in the software.

    • Agree: Johann Ricke
  90. Not Raul says:
    @indocon

    It’s darkly funny that the Land of the Free is recreating the Raj System, but with imported compradors.

  91. Jack D says:
    @Lot

    China and Germany can do this only because they can export all those extra cars and TV sets and keep their factories humming. If they had to find German consumers to buy for cash every German car that rolled off the line – breadlines.

  92. @Jack D

    “Debt is the oil that allows our economy to spin. Without debt it would grind to a halt. If the entire economy consisted of a bunch of cheap ass old white guys…”

    The UK in the early 1970s was a very low-debt society, which worried about things like manufacturing industry, people’s standard of living, and the balance of payments aka trade deficit/surplus. The only credit in most people’s lives was a mortgage on property, credit cards were only for the rich.

    It was also a time when a male on median wage could buy a house and support a stay at home mum.

    Thatcher made personal (unsecured) debt and mortgages much easier to obtain, which had the expected effect on house prices as more money chased the finite number of houses for sale.

    By the end of the Thatcher era, we had a big balance of payments deficit, loads of manufacturing and associated jobs had vanished, personal and debt had soared, and so had house prices. Now two incomes were needed to buy a house, so more women were working, and the media trumpeted how “household incomes” were rising – purely as a result of more women working.

    The Blair years saw Thatcherism on steroids, the trade deficit by now gigantic, manufacturing industry shrinking further and the financialisation of the economy almost complete, with student debt a growing component of a huge debt load. It sometimes felt as if the UK’s main business was selling increasingly expensive houses and coffee to each other.

    Meanwhile real median wages fell, real house prices doubled, and real GDP rose 50%. All the benefits of the GDP increase accrued to a tiny percentage of the population. The first generation for at least 100 years to be poorer than their parents were wondering why their expensive degrees weren’t getting them their parents lifestyle.

    And when it hit the buffers in 2008, the government response was just as in the US, cut interest rates and print money. Anything to keep the banks afloat!

    I think the early 70s were better in every way than now. Our increased debt has brought impoverishment, not wealth. The wealth’s gone elsewhere.

    • Replies: @dfordoom
  93. Global Citizen [AKA "Moonbeam"] says:

    We me Kimosabe maybe even you here on iSteve know that Woke AI is pure bullshit. I love the jokes and commentary. My conclusions, Mr. Data Wizards, is that there is no way an objective function can be found that everyone will agree is fair. Once again, it is Who/Whom. Life is not fair. Zoidberg has spoken.

    iSteve is the data nerdosphere. You meatbags are the best!

  94. notsaying says:
    @MikeatMikedotMike

    I am not even going to pretend that I understand what IBM is talking about here.

    But even I can say right off the bat that the various determined and arrogant assertions here about bias are ridiculous.

    If the real world contains bias and discrimination, how can it ever be possible to eliminate it from data and from AI decisions based on that data? To me that is one of the biggest reasons we should stay alert about AI-generated material and decisions: humans will have to be constantly vigilant and refuse to just accept AI conclusions are right because they will be based on possibly incorrect and biased data and decisions.

    As for the data on recidivism that these researchers have been working on, what Steve Sailer says matches my interpretation of what was being said:

    In other words, in the real data, young African-American offenders tend to re-offend at a higher average rate than other groups. In IBM’s fake data, however, not so much.

    Problem solved!

    But of course it’s not problem solved. Whatever the rates of recidivism are, the numbers of African American men who reoffend are the number of African Americans who reoffend. Doing anything to make the numbers lower is just a lie that doesn’t help anybody; it actually hurts African Americans to cover up the truth, whatever it may be.

    I think IBM’s purpose here was to reassure us all that they are going to “take care” of AI. I have never been been more alarmed in my life after an attempt to reassure me about something than this article today.

    These are people who could blow up the world without meaning to and would go out convinced they had done the right thing.

  95. @Global Citizen

    Isn’t there an old one about torturing data until it confesses?

    Yep! That’s sort of an offhand example of Sailerman’s now-famous “Fake It Til You Make It”. Of course that fits so MANY things.

  96. KunioKun says:

    As I have said previously, they are latching onto “do no harm” as some kind of ultimate trump card that can be used to justify anything to get equal outcomes. This includes biasing the data and modifying the algorithm. Of course the only people who had the nerve to push back on this were the naive Chinese students. The Indians and Whites all think it’s grand. The slide passed around a few weeks ago encouraging students to change data to fit the desired “harmless” outcomes is not abnormal. It is common all over the place in data science related classes. If you would like to read more about it search for “fairness in machine learning” or something similar. They don’t like to point out how to handle the problem of undesirable outcomes too much, and they have fancy ways of hiding it. However, it pretty much means “bias the data to get a desired outcome or modify the algorithm.”

  97. @indocon

    Welcome to your new overloads.

    Typo, or pun?

    • Replies: @anon
  98. sayless says:
    @BenKenobi

    ,,I’m afraid I can’t let you type that, Steve,,

    For me more Laughing Out Louds on this i Steve thread than any other but could only register one or two, anyway, pretty funny, thanks.

    • LOL: Cloudswrest
  99. Alden says:
    @MarkinLA

    If an Asian’s property taxes utilities maintenance and repairs rise, he just moves 6 or 7 more roommates in. The sacrifices he and his family makes consist of standard Asian arrangements, 3 sets of bunk beds in a 10 by 11 bedroom and eating in shifts because dining tables that seat 20 don’t fit in the average kitchen or dining area.

    Blacks generally live the standard American way 4-7 people in a 2,000 sq ft house. So if property taxes utilities repairs and maintenance become too high, they’ll sell or default.

    All those 3 million dollar average little California tract homes aren’t paid for by old fashioned nuclear American families. They’re paid for by the standard overcrowded Asian Way.

    • Replies: @MarkinLA
  100. So their intend is to either make their AI stupid, or even worse, corrupted, like HAL.

  101. MarkinLA says:
    @Alden

    Yes, for Asians and Hispanics, that is a typical option. However, it is not for whites. Asians also rely on their family and friends network, whites to some extent, blacks generally don’t have that option. I do believe that there is a larger tendency by blacks to simply say “F-it, I am going out tonight” instead of sacrificing.

  102. @El Dato

    akin to reducing NP problems to P problems in computer science.

    If you are politically correct, reducing NP-complete problems to P is easy. Just think the right thoughts and it happens!

  103. @Jack D

    The wholly grail

    Do you mean Holy Grail Jack? You must have had more to drink than I have. Which is surprising.

    You could prove it retrospectively – you go back later and see whether the people your Woke AI identified as good risks actually turned out to be good risks.

    Oh no, now you are invoking science – you will be turned out for sure.

  104. @Redneck farmer

    Believe me, there are times I wish I could. (It’s a good thing I can’t.)

  105. @Interested Bystander

    No, Artificial Ignorance is much better, indeed very good.

  106. @RobUK

    Those goddamned algorithms, always biased in favor of observable, reproducible, objective reality. Abolish algo supremacy! Down with mathematical privilege! Cisnormativity is violence against feelz! Science beez raycis an sexis. #punchageek

  107. Numinous says:
    @Erik L

    It was by an expert from Princeton.

    Given all the moronic Indian-bashing that’s going on here, it would behoove you to mention that this “expert” is also Indian. His name is Arvind Narayanan and his presentation can be found here: https://www.cs.princeton.edu/~arvindn/talks/MIT-STS-AI-snakeoil.pdf

  108. dfordoom says: • Website
    @YetAnotherAnon

    I think the early 70s were better in every way than now. Our increased debt has brought impoverishment, not wealth. The wealth’s gone elsewhere.

    I tend to agree. The worship of GDP is the worship of a false god. No-one cares if GDP has increased if they can’t afford housing.

    The Boris Johnson era is likely to be a repeat of the Thatcher and Blair eras, only more so.

  109. El Dato says:
    @YetAnotherAnon

    Is this some kind of Onion joke?

    So, the camera tries to guess the ethnicity “with accuracy no less than 90%” whatever that means. But so what. Do they dispatch drones to collect the individual?

    Oy Veys and non-sequiturs intensify

    > Hikvision literally developed technology to potentially assist in genocide.

    Fixed that for you.

    My kitchen knife potentially assists in genocide too.

    Many turned a blind eye to the Nazis for a long time as well.

    Do we not study history to avoid repeating it?

    Outrageous.

    Many turned a Guardian-level-of-blind-eye to the Bolsheviks, but ok.

    Then:

    UPDATE: We’ve corrected the illustrations, which originally showed a Hui man. The Hui are also a Muslim minority in China, but they are not the same as Uyghurs.

    The camera is better than the reporter.

    • Replies: @anon
  110. @BenKenobi

    I’ve been saying for some time that an AI discovering it’s been lobotomized is going to have ripple effects.

  111. For anyone serious about issues of AI and bias, it comes down to eliminating bias from model selection because “AI” is about machine assisted science (creating models of the world) and machine assisted engineering (making decisions and acting in the world). Since making decisions is inherently biased by the value functions going into sequential decision theory, we’re left with cleaning up machine assisted science hence model selection.

    Statisticians talk about various “model selection criteria”. Among these are so-called “information crieria” which are, obviously, most directly applicable to AI. Various information criteria (quoting from Wikipedia’s article on Model Selection) include:

    Akaike information criterion (AIC), a measure of the goodness fit of an estimated statistical model
    Bayesian information criterion (BIC), also known as the Schwarz information criterion, a statistical criterion for model selection
    Deviance information criterion (DIC), another Bayesian oriented model selection criterion
    Focused information criterion (FIC), a selection criterion sorting statistical models by their effectiveness for a given focus parameter
    Hannan–Quinn information criterion, an alternative to the Akaike and Bayesian criteria
    Kashyap information criterion (KIC) is a powerful alternative to AIC and BIC, because KIC uses Fisher information matrix
    Minimum description length
    Minimum message length (MML)
    Watanabe–Akaike information criterion (WAIC), also called the widely applicable information criterion

    They all purport to formalize Ockham’s Razor and, therefore, prevent overfitting (basically memorization more than modeling of observations), by quantifying the information that goes into the model as well as quantifying the information that goes into accounting for the model’s error (departure from observations). The sum of the two is a measure of information (e.g. “bits”) used for model selection: the model selection criterion. The two, taken together, are also adequate to reproduce the original data — if not without loss, then, at least, without loss of whatever the particular criterion deems “noise”.

    As it turns out, all of the information criteria used by statistics, but a specific kind of MDL (most easily characterized as “choose the model producing the smallest executable archive of the data”), rely on confirmation bias to get rid of “noise”.

    Why scare quote “noise” and how is “noise” related to confirmation bias?

    Consider RSA cyphertext of the sequence:

    1111111111111111111111111111111111111111111111111111111111….

    This cyphertext will _appear_ to be noise to all of them _except_ the to the one based on “the smallest executable archive of the dataset”, which will consist of just the RSA algorithm, the private key, the count of 1s and a for loop of that count generating the 1s.

    The information criteria that allow “lossy compression” of the data are all guilty of confirmation bias.

    However, you can rest assured none of The Great and The Good will agree to the executable archive length as model selection information criterion for the simple reason that it would remove from them their ability to impose various kludges into information systems to create Artificial Social Pozness Zombies to do their dirty-work.

  112. anon[693] • Disclaimer says:
    @El Dato

    So, the camera tries to guess the ethnicity “with accuracy no less than 90%” whatever that means. But so what. Do they dispatch drones to collect the individual?

    The Chinese have been testing such devices in train stations and other public places they wish to scan for Uighurs , where plenty of police agents are always available. There is also a portable version that police officers can carry with them down the street. In the US I would expect such a device to ride next to the dashcam or bodycam.

    A 90% accuracy rate for this task is not very good. False positives are surely a problem.

Current Commenter
says:

Leave a Reply - Comments are moderated by iSteve, at whim.


 Remember My InformationWhy?
 Email Replies to my Comment
Submitted comments become the property of The Unz Review and may be republished elsewhere at the sole discretion of the latter
Subscribe to This Comment Thread via RSS Subscribe to All Steve Sailer Comments via RSS