Comment by paulgb
5 years ago
In a similar vein, by the same author: https://www.mathwashing.com/
(It's linked at the bottom of this one, but I'm sure a lot of people don't get that far)
5 years ago
In a similar vein, by the same author: https://www.mathwashing.com/
(It's linked at the bottom of this one, but I'm sure a lot of people don't get that far)
That is my number one pissing point right now in higher education.
Every company has a predictive algorithm to use on students. Every startup that's stepping into the space is pushing the data and data scientists.
But they all have the same-old, usually decades old, baked in biases. AND they're not doing anything to address it!
Just because it's math doesn't mean it's not biased. I hate it more than anything professionally, right now.
> Just because it's math doesn't mean it's not biased.
When a model produces an unpalatable result, that doesn't mean it is biased. All these algorithmic fairness people are saying, once you peel back the layers of rhetorical obfuscation, is that we should make ML models lie. Lying helps nobody in the long run.
>When a model produces an unpalatable result, that doesn't mean it is biased.
Absolutely, but there is no admission, from what I can tell, from ML or predictive [buzzword here] companies that bias is a thing, or even could be a thing in their systems.
>All these algorithmic fairness people are saying, once you peel back the layers of rhetorical obfuscation, is that we should make ML models lie. Lying helps nobody in the long run.
Maybe I'm misunderstanding, but that is not at all what I am saying as an 'algorithmic fairness' person. I am saying that we need to ensure there are strict oversights and controls on the building/execution of algorithms when making substantive decisions about human people.
For example: It's okay if an algorithm predicting student success says that all the minority students on my campus are at a higher risk of dropping out. That is a data point. Historically, minority students drop out at a higher rate. Sure. Not great, but it is factually true.
What is not okay is for the 'predictive analytics' company to sell their product in conjunction with a 'tracking' product that limits minority students' access to selective admissions programs simply because they are selective, more difficult, and, historically, have a higher percent of minority students who drop out.
I guess what I'm saying is that ML models shouldn't lie. But they also shouldn't be seen as the truth above all truths. Because they're not. They're just data, interpreted through the lens of whoever built the models.
Every human carries a bias, everyone. It's how we define ourselves as 'self' and others as 'other' at a basic level.
Therefore, everything we build, especially when it's meant to be intuitive, may carry those biases forward.
I'm only saying we need to be aware of that, acknowledge it, and ensure there are appropriate controls and oversight to ensure the biases aren't exasperated inappropriately.
What do you think about the second issue : the "blackboxing" ?
Oh wow, this is so juvenile.
The problem with these ML models is that they are directly connected to a response, and because they use statistics and math they are simplistically perceived as truth. They're not truth, they're are nothing more than models. The truth can fit a plethora of models. Ignoring human bias while training ML models, and then just saying that the model is truth is exactly the problem.
Thank you for demonstrating the issue so vividly.
1 reply →
It takes years and years of training in advanced dialectic bullshit to get to the point where you can say, with a straight face, that math is morally wrong. It's utterly absurd to demand that we censor models and worsen their output to conform to some activist's idealized imagine of how the world should be. Only by letting models report the true facts of the world as it is can we optimize for human happiness.
Math is a language for modelling things. It can be intrinsically correct, as in consistent, but that doesn't say anything about the model's actually validity.
Every choice we make is a moral choice. Once we're done modelling and use that model then we make a moral choice.
For example, If you believe that lowering the debt default rates is more important than the fairness to an individual.
Then you make a moral choice. Of you believe it is OK to not give loans to Blacks because there's a largish amount of Blacks defaulting on their loans thats a moral choice.
Further more, Enscribnng truth to models is just an age old human fallacy. The truth can somewhat fit plenty of models. None of the models are truth.
It's not the math, that is wrong. The math is correct.
The inputs and assumptions made by the people selecting the math is the 'morally wrong' part.
Bias is real, like it or not. Your worldview, as a data scientist or programmer or whatever, impacts what you select as important factors in 'algorithm a'. Algorithm a then selects those factors for other people in the system, baking in your biases, but screening them behind math.
That's the motte. The bailey is that the ML fairness people use any inconvenient output of a model as prima facie evidence that the model inputs are tainted by bias --- then these activists demand that these inputs be adjusted so as to produce the outputs that please them. They've determined the conclusion they want to see beforehand. This attitude is the total opposite of truth seeking.
1 reply →