A Collection of Tutorial Comments

Preface

As someone primarily interested in technology philosophy I’ve been fortunate enough to have functionally taken 2 classes with Prof. Johnson on the subject, so much of the papers came as review to me. In rereading I had more thoughts come up about how they presented the argument then the argument itself. Thus, I thought it would be more interesting to write a short pseudo tutorial comments paper in reaction to several of my peers blog posts. Also since last time some people suggested I just paste entire relevant past papers I’ll do add one at the bottom. I will sometimes pull quotes from it in commenting.

Tim on Weighting, Internet Training, Bias in the data, and Human vs AI bias

I’m glad you brought up weighting because it’s the intuitive CS answer. However, it really would not solve the problem.

“These are simple examples of data manipulation, the extent to which you can shift weighting of features in a data set is near limitless. Parts of data sets can easily be given x times as much weight, enabling flexible results. Additional algorithms could even be developed to automatically adjust the weighting to maintain equivalence. For instance, decreasing the weighting of women data points if the number of searches on “women in stem” increases, or increasing weighting upon decreases. The flexibility, simplicity, and speed of these corrections outline the effectiveness of debiasing through data manipulation. Hence, the adoption of these strategies over the past few years.

Next, I challenge debiasing data manipulation techniques effectiveness in creating non-problematic data. I contend there are problems with relying on solely debiasing data manipulation for mitigating problematic social biases. These debiasing techniques privilege subsets in ways that reproduces bias through other characteristics. Existing women in stem may have many characteristics not present in women generally such as higher parental wealth. As such a debiased algorithm identifying future stem majors may identify an equal number of men and women, but it would be disproportionately weighted towards the affluent. This means it does not resolve the fundamental problem of reproducing past trends. Some women who do not grow up in affluent families should become stem majors. But if most past successful examples are of affluent women, algorithms will identify the non-wealthy as unfit. There may be an attempt to forcibly decrease the weighting of wealthy stem majors. Then the algorithm will be dominated by men stem majors who are far more common among middle- and lower-class upbringings. Additionally, reweighting the number of data points does not change percentages. Multiply the number of white defendants in Northpointe’s algorithm. Those white defendants’ false flag rate remains the same, not equivalent to black defendants, since it simply increases the size of the existing data set. This reproduces an algorithm that is discriminatory in how often it concludes white defendants will reoffend. Third, the “black box” nature of many complex algorithms means engineers cannot explain, let alone effectively modify these programs while maintaining accurate predictions. Fourth, bias comes from what is absent from the data set. People who search about “women in stem” have technological access, many women who are interested may not. If two women are considering stem. A heavy tech user will likely search on google, while the other women may ask a friend. The light tech user will have their biases underrepresented in search datasets. These biases are nearly impossible to manipulate due to creating absences within the training data. These doubts challenge debiasing data manipulation techniques’ ability to prevent individual algorithms from inheriting problematic social patterns.”

Additionally, a majority of algorithms do not have open access to the internet, prominent examples like Microsoft’s Tay in the paper have taught some lessons, they are mostly taught offline. The problem is the same human biases exist offline as on the internet.

“What an algorithm does is that it seeks to produce the most correct result with the data it is presented. This is not unlike what a human (such as a judge) does - we try to make the best decision with the data that are presented.”

“most correct” is an interesting phrasing, and I think would be something we ought to explore in seminar: what is the most correct result to strive for. Because the problem is often how the algorithm defines that.

“an algorithm's bias is actually capable of being corrected and improved, whereas our biases are static.”

This seems false, humans are famously distinct from other animals on our ability to learn. Also seems to lead down a problematic road of racist doomerism, where a racist is permitted to continue to be racist because you can’t teach an old dog new tricks. It seems just like AI we can work to reduce human biases.

“When we see that the algorithm unfairly punishes African Americans, we can easily attribute that mistake to a misevaluation of the data that fed its systems.”

First the algorithm did not just do bad compared to perfect, it did significantly worse then humans.

Second, this really misses the point Johnson is trying to make, which is the Computer Scientist always argues the bias is tied to the data and if they had better data then they wouldn’t have bias. This sort of “view from nowhere” ignores how the data becomes biased in the first place.

“Beyond accounting for algorithmic bias resulting from data inputs influenced by social structures, I contend we must account for bias in the neutrality heuristic used surrounding algorithmic decision-making. The neutrality heuristic inhibits effective countermeasures to bias in data under the pretense that it is already neutral. It propagates a myth of objectivity, of no algorithmic decision. If you asked Trump to set the top ten search results for the word “spy,” you could scrutinize their choices. However, when google selects what results to display, it is easy to forget a decision was made at all. Yet, on google two people can get different search results with the same term. Bias does not appear in data from nowhere. Bias originates from a broader social context effecting data, and a heuristic of neutrality that pretends its nonexistent. Relegating bias to only the data supports the myth that future debiasing data manipulation will produce the perfect algorithm. It leaves unexamined decisions with broader social context. What elements do algorithms choose to examine? Where do they source their data from? Debiasing data manipulation can provide easily implementable short-term assistance. But, heralding neutrality prevents developing long-lasting solutions to problematic structural social patterns in algorithmic data and decision-making paradigms.”

But yeah Humans are more biased then AI in some areas. I think it should be evaluated case by case which should be used, but this idea that AI can be perfect and humans can’t is leading to some really discriminatory practices.

Kara on Economic Rationality

“Here we make assumptions about the agent. They are forward looking (meaning they care about their future). Ex. if you’re working as a gov official and get your check, you’re going to save some of that money for the future and not consume all of it now. They desire to smooth consumption across time. This means they want their consumption in this period and next period to be smoother, where they aren’t spending it all today or tomorrow. Further, current consumption should not reach much to predictable changes in income that were predicted in the past. Lastly, these agents are rational meaning their behavior is optimizing, so they are optimizing utility (satisfaction, happiness). I will only focus on the assumptions of this model, to keep it simple.”

Two quick questions 1. I’ve never understood what we rely on to make these assumptions about the agent, most agents I know are not fiscally responsible or significantly forward looking, or can’t be. Does that justify an economic theory where its economically rational to spend all your money on recreation?

2. I think there’s interesting interaction with Prof. Johnson’s presentation of Hume’s problem of induction here. Since we have no logical reason to believe past evidence supports how the world will operate in the future, isn’t it illogical to be future oriented?

Rukmini on Textualism

Yeah I think I brought up Johnson last semester in Phil of Law to make a different argument, and I was kinda of learning and understanding the canons of inference through the epistemic values lens. I’m glad to see someone else making connections between these phil of science arguments and legal interpretation!

Nicole on value-laden decisionmaking

“Recognizing this reality is crucial. We cannot expect individuals in the public or even judges who work in the criminal justice system to stop trying to apply the probabilistic findings of scientists or algorithms to real life situations. The important question, then, is this: who is the person who is best suited to be making the value-laden judgements that must be made? “

I think these two sentences nail it with getting to the crux of the question. Just to add to the paper as a whole, regardless of who should be, the scientist will definitely not be the decisionmaker given their fear of value-laden judgements, leaving the options really to the judge and the algorithm. As you say, judges will be influenced by the algorithm, but a distributing trend that compounds this is court packing and our legal systems heavy load on justices for judicial processing. Judges time per case only appears to be getting shorter and shorter so its no surprise algorithms were rapidly adopted for quick decision-making. Just like with plea deals I fear we will favor the faster option, then the more just one.

Old Paper to make Hurley’s printing life more difficult. I wonder if each Blog post I write is longer than the last where we will be at the end of the semester? 50 pages? For comparison this one is 19. Comment your guess.

The Inductive Failings of Debiasing Data Manipulation, Towards a Structural Approach to Algorithmic Bias in Data and the Heuristic of Neutrality

First, I explain how Induction’s risk of a false conclusion means algorithmic decision-making inevitably incorporates bias. I explore how algorithmic biases arise from data. I present debiasing data manipulation techniques to correct biased data. Then I challenge debiasing’s ability to mitigate social biases effect on data. I outline the failure of individual program bias mitigation strategies which do not change associated structural patterns. I use induction to illustrate how algorithms reproduce the problematic social environment of their training data. Finally, I explore algorithmic bias in the heuristic of neutrality surrounding algorithmic decision-making. I conclude attempts to eliminate algorithmic bias through solely debiasing data manipulation is ineffective.

Inductive risk means algorithmic decision-making contains bias. An inductive argument’s premises’ truth does not guarantee the conclusion’s truth. Induction is used to bridge the inductive gap between evidence and a potential conclusion. We may observe billions of exclusively black ravens. Yet, to conclude all ravens are black, including unobserved ravens, requires something beyond past evidence. This additional value must warrant that the evidence is sufficiently strong or probable to overcome the inductive gap and accept the conclusion. This supports premise 1: every induction risks getting things wrong. There is always a potential unobserved grey raven. Premise 2: since every induction risks being false, inductive decision-makers must decide at what point their inductive support is sufficiently strong to take on that risk. An algorithm that tests drugs, must reach a point where it decides some drug is safe enough to release for use. Different rational actors may decide at different thresholds. An algorithm may choose to release the drug at 70% safety, viewing the potential benefits of immediate deployment as worth the risk. Another more risk-averse algorithm or scientist may wait until 99% safety for the same drug. There will never be a 100% guarantee of safety. In making the decision to release a drug, some values, which contain bias, are necessary to cross the inductive gap. The conclusion is bias is a necessary component of induction. Just as scientists carry inductive biases into the lab, no algorithm is perfectly neutral. Not all biases are inherently problematic, and there is no shame in admitting algorithms possess them. However, there are numerous examples of algorithms possessing problematic social bias. Google autocorrected to “himself” rather than “herself” upon receiving “English major who taught herself calculus.” [1] (Noble 2018, pg. 142) In this case, Google’s search algorithm decides a typing mistake for a man is more likely than a woman intentionally typing the phrase.

Algorithmic biases arise from problems in the data. In 2016 ProPublica found Northpointe’s risk-assessment algorithm for criminal sentencing racially biased. It was nearly twice as likely to falsely flag black defendants for future crime as whites.[2] (ProPublica 2016) White defendants were falsely labeled as low risk for repeat crime more often than black defendants. Northpointe’s programmers did not explicitly write racist lines of code. In fact, from the 137 questions the algorithm uses, race is not included. Thus, even without the explicit use of social factors in algorithmic decision-making, algorithmic decisions are influenced by social patterns. Rather than programmed biases, algorithms develop biases from patterns in the data upon which they are trained. For example, an algorithm receives data about people being arrested in a city. Of the arrests 90% are black. The algorithm will conclude that black people in that city are likely criminals. It will not account for the high police presence in black neighborhoods creating high arrest percentages. Even if it avoids race as a factor, secondary factors such as socio-economic status and cultural touchpoints will reproduce similar effects along racial lines as displayed by ProPublica. Since algorithms learn from patterns in past data, they will reproduce those same patterns. This means if existing data is problematic and socially laden, inferences made by algorithms from it will be as well. Therefore, to create algorithms which do not reproduce existing problematic biases we need to fix their data.

Now, I provide debiasing data manipulation techniques to correct for problematic biases in algorithms’ data. A popular debiasing approach is to amplify the representation of an underrepresented group. Google’s algorithm autocorrects “women in stem” to “men in stem.” Google can multiply the number of “women in stem” search cases until it equals the number of searches of “men in stem.” This approach of amplifying representation by multiplying existing counter-stereotypical examples can help prevent existing social stereotypes from reproducing themselves. With this technique women will be shown similar search results to men, and not pushed into stereotypical gender roles by algorithms. If applied to Northpointe, white defendants would no longer be underrepresented in the data. These are simple examples of data manipulation, the extent to which you can shift weighting of features in a data set is near limitless. Parts of data sets can easily be given x times as much weight, enabling flexible results. Additional algorithms could even be developed to automatically adjust the weighting to maintain equivalence. For instance, decreasing the weighting of women data points if the number of searches on “women in stem” increases, or increasing weighting upon decreases. The flexibility, simplicity, and speed of these corrections outline the effectiveness of debiasing through data manipulation. Hence, the adoption of these strategies over the past few years.

Changes to mitigate algorithmic bias cannot happen at the level of individual programs and must account for structural patterns. A major problem facing debiasing data manipulation techniques is their inability to resolve social structures’ effects through implicit biases. A definition of implicit bias resulting from social structures is “simply ideology that is internalized while occupying the social structures it supports.”[3] (Haslanger 2015, pg. 8) I contend we recognize the existence of implicit algorithmic biases to challenge problematic biases. If past examples were insufficient support, I propose implicit biases are necessary to algorithms. Algorithms only have value insofar as they can explain for a person’s behavior beyond explicitly stated characteristics. The unbiased algorithm is unable to make inductive inferences and needs the input stem major to determine a person is a stem major. It does not produce any knowledge. In mapping implicit bias, each data node depends upon background data points. Individual characteristics should not be viewed in isolation, but as a web of associations. Failing to mitigate bias in associated characteristics results in bias reproducing itself even when explicitly corrected for in individual connections such as between women and stem. Manipulating the data of individual programs cannot resolve the larger surrounding web of social biases that construct the data. A perfectly debiased racially balanced crime recidivism algorithm sounds wonderful until you realize the sentencing algorithm gives five times the prison time too black defendants. This does not mean you have one bad program and one good program. Rather it reveals the crime recidivism algorithm is predicting black recidivism at five times the rate given their higher sentences. Therefore, while external solutions to problems in the data are necessary, they must go beyond fixing immediate inputs of an algorithm.

So long as problematic social patterns are present in the environment, the data which arises from that environment will train an algorithm which reproduces those problematic social patterns. When we understand algorithmic decision-making through the lens of induction this conclusion becomes predictable. When deciding on a conclusion from a pool of evidence, characteristics, or premises, the algorithm employs the results of its training data to overcome the inductive gap and decide. This means the bias used to decide a person is sufficiently likely to become a stem major is based on past examples. Therefore, all instances of algorithmic induction will always make decisions regarding new cases by closely reproducing its past data set. That is not inherently bad, but a constitutive component of algorithms that prevents them from anything beyond predictions which recreate past. Therefore, since we live in a world which is thoroughly influenced by problematic social patterns, if we want algorithms to not recreate those patterns with corrupted data sets, we need to change those data sets externally not solely via code.

Beyond accounting for algorithmic bias resulting from data inputs influenced by social structures, I contend we must account for bias in the neutrality heuristic used surrounding algorithmic decision-making. The neutrality heuristic inhibits effective countermeasures to bias in data under the pretense that it is already neutral. It propagates a myth of objectivity, of no algorithmic decision. If you asked Trump to set the top ten search results for the word “spy,” you could scrutinize their choices. However, when google selects what results to display, it is easy to forget a decision was made at all. Yet, on google two people can get different search results with the same term. Bias does not appear in data from nowhere. Bias originates from a broader social context effecting data, and a heuristic of neutrality that pretends its nonexistent. Relegating bias to only the data supports the myth that future debiasing data manipulation will produce the perfect algorithm. It leaves unexamined decisions with broader social context. What elements do algorithms choose to examine? Where do they source their data from? Debiasing data manipulation can provide easily implementable short-term assistance. But, heralding neutrality prevents developing long-lasting solutions to problematic structural social patterns in algorithmic data and decision-making paradigms.

In this paper I have explored how debiasing data manipulation techniques are an ineffective long-term solution to algorithmic bias. The process of induction means inevitably algorithmic decision-making needs bias to make decisions. Problematic biases arise from both algorithms’ training data and decision-making heuristics. Debiasing data manipulation techniques fail because they overly isolate individual cases, ignore data’s web of associations, and avoid explicit confrontation with bias due to neutral heuristics. Algorithms will always reproduce patterns in past data so solely algorithmic solutions can not rectify problematic patterns. Therefore, broad external structural changes to society and heuristics are necessary to develop long-term solutions to the problem of algorithmic bias.

Bibliography

“Algorithms of Oppression: How Search Engines Reinforce Racism Introduction and Ch. 5.” Algorithms of Oppression: How Search Engines Reinforce Racism, by Safiya Umoja Noble, New York University Press, 2018.

Julia Angwin, Jeff Larson. “Machine Bias.” ProPublica, 23 May 2016, www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing.

Distinguished Lecture: Social structure, narrative and explanation Sally Haslanger Canadian Journal of Philosophy, 2015 Vol. 45, No. 1, 1–15, http://dx.doi.org/10.1080/00455091.2015.1019176

[1] (Noble 2018, pg. 142)

[2] (ProPublica 2016)

[3] (Haslanger 2015, pg. 8)

Search This Blog

PPECMC2022Sextons