[jiofih]

Same feeling here - instead of laughing I was disappointed when the second paragraph laid out the April fools joke.

I would love to see “small batch data science” where people actually understand the results. Black box recommendations never feel really right.

[disgruntledphd2]

It's called statistics, to be fair.

In general, if your DS's can't explain the model then something has gone horribly wrong.

[fxtentacle]

People usually just call it AI then and pretend that it's not a problem.

For example, there's been plenty of issues like racial bias with the computer vision algorithms that police uses, which effectively is data science on pictures. But nobody knows why the issue occurs and nobody can specifically fix it without the risk of breaking a thousand unrelated other things.

[shakna]

> But nobody knows why the issue occurs

Sure we do. Photography itself has a racial bias. [0] Different skin tones produce different levels of detail, and it has been an uphill battle to be able to capture those details since the advent of photography.

So long as facial recognition relies on photography, and photography is flawed, then every dataset is biased. Which will exacerbate the bias of the AI, which already has its own set of problems from where those datasets are formed.

[0] https://www.nytimes.com/2019/04/25/lens/sarah-lewis-racial-b...

[ZephyrBlu]

> But nobody knows why the issue occurs

They do know why that occurs. It's because the data set is biased.

[tremon]

No, you don't "know" that your dataset is biased until you perform the statistic analysis explicitly. It might be that your neural net has a non-uniform weight distribution in some dimension (e.g. in time, or in the ordering of the training data), so dismissing any unwanted results by claiming "your dataset is biased" is a form of appeal to (artificial) authority.

[marcinzm]

It's not an appeal to artificial authority. It's a very likely root cause and comes with a solution even: get a different data set or adjust your existing data set. Your response rings of No true Scotsman to me since you can argue any analysis is not rigorous enough or doesn't cover all potential issues.

edit: And a statistical analysis isn't some sort of magic data genie. Statistics can give rigorous results because it makes strong assumptions. If those assumptions don't hold then the results aren't rigorous anymore. A trillion parameters model can pull interactions out of your data that almost no statistical analysis of the data would identify ahead of time. So what you need to analyze is the model and try to infer why it's predicting different certain results and then work backwards from there.

[jokethrowaway]

Is it a case of BAME have less stable families, are poorer, commit more crimes and therefore are more represented in the data which lead to even more incarcerations?

[mauvehaus]

First: I don't think you can claim that without also doing some very rigorous statistics. I'm not asking you to, but if you're going to base policy on that statement rather than merely arguing on the internet, you'd need to.

Second: even if you do, you're going to have a hard time controlling for the fact that the police and criminal justice system has a long history of disproportionately enforcing the law against people of color. The base data about who commits crimes, gets convicted, etc. for well over a hundred years is going to reflect this bias.

I'm not claiming to have done my homework here either, same disclaimer applies. I suspect you could find somebody who does study this if you wanted to look.

All of the above assumes we're discussing the US, btw.

[alexashka]

Funny, I remember thinking the questions were biased, every time they weren't the ones I studied to answer in my exam preparations.

Too bad I wasn't a data scientist or else I could just get a passing grade by claiming the questions were chosen from a biased data set, or retake the exam until the data set matched the questions I studied for, at which point the data set would no longer be biased, lol.

Funny line of work, this data 'science' where you only use the results that fit the narrative you wanted in the first place.

We're in full doublethink mode, just keep repeating data 'science', 'science', 'science'. :)

[jjgreen]

I did a geography O-level (the UK exams for 16-year olds at the time) which included a map-reading exercise. It just so happened to be a couple of miles from where I lived, and I knew it well. Still only got a B though.

[marcinzm]

>In general, if your DS's can't explain the model then something has gone horribly wrong.

With modern ML models data scientists can generally explain why the model should work but not exactly how it gives any specific prediction. Explaining how a model with a trillion parameters comes up with an answer is not a trivial task.

[disgruntledphd2]

> modern ML models

As an aside, this is a hilarious phrasing. What are we going to call these methods in a decade? You'd probably be better off phrasing it as deep neural networks.

Anyway, this is not really true. There are methods (for lots see: https://christophm.github.io/interpretable-ml-book/) and the DeepDream paper came out in 2015, so it's possible.

It's computationally expensive, and a lot of people don't see the value, but my argument is that if you want to use the model in the real world, and have non-technical stakeholders then you'll need to do this, and in general, I've found it to be the best way to actually improve a model.

And to be fair, if you just need to see how the predictions vary as a function of the inputs, you can again hold all but one constant and run a bunch of different values through the model.

Again, this can all be done, but I think it's more a question of will than capability (and hence the context of my original commment).

[marcinzm]

>You'd probably be better off phrasing it as deep neural networks.

I phrased it the way I did for a reason. Large Random Forrest models are also not easily interpretable. Even large logistic regression models with feature interactions and feature hashing aren't easily interpretable. It's not a question of the model technique used but rather the amount of parameters and how many feature interactions are modeled.

>And to be fair, if you just need to see how the predictions vary as a function of the inputs, you can again hold all but one constant and run a bunch of different values through the model.

This only provides a partial view of the feature impact due to non-linear interactions.

[DoreenMichele]

I have been working on a spread sheet on and off for a few weeks. The data exists, but you can't google the answer(s) I want, so I have been painstakingly putting it together, a few minutes here and a few minutes there.

I've been calling it my "curiosity killed the cat" spreadsheet. Small batch data science sounds so much more respectable while meaning basically the same thing, I think.