The Off Topic Topic

[Grey Fox]

What the hell is this?

Far right wing weirdo in US wants to make a truly true series about the Celtic Celtics (not Celts!) to show how they really were. Mostly seems to contain "muh speech!" and "purity of race" stuff.

How can you do an entire series about the Celtics without a single reference to Larry Bird?

With constantly applied restraint.

[The Larch]

I guess there's no point in explaining this dude that the Roman Empire never had anything to do with Ireland, right?

[Josquius]

To these people Scotland and Ireland are interchangeable. And the Romans totally stopped at the Scottish border cos freedommmmn

[Maximus]

Reading a story about Facebook and online ads. It said about how all of us have had the weird or embarrassing algorithm stories - but then had the story of Melissa Edwards, 38, from London who lost her twins for months into her pregnancy. She continued to receive maternity-related ads for months and said she basically ended up repeatedly googling "miscarriage" in the hope that it would sort of shift the algorithm away from constantly showing her baby ads.

Which is just incredibly sad - an accidental cruelty built into someone's life every day

As I understand it, that's a pretty common complaint. It's certainly something I've heard (in media) about more than once over the last many years.

Data science is a brave new world.  There is so much potential in it, but also so much danger due to both widespread incompetence as well as lack of ethical guidelines.

I imagine it's not helped by the demographics of the teams creating the models and data sets.

Also negation is actually pretty hard in NLP.

[crazy canuck]

To these people Scotland and Ireland are interchangeable. And the Romans totally stopped at the Scottish border cos freedommmmn

Mel should do a movie about that

[DGuller]

I imagine it's not helped by the demographics of the teams creating the models and data sets.

Also negation is actually pretty hard in NLP.

It's even less helpful to make problems with data science be about the demographics of data scientists.  Not only is it offensive, but it's also wide off the mark.

[Oexmelin]

Why is it wide off the mark?

[DGuller]

Why is it wide off the mark?

Because it's a lazy template answer that can be (and is) applied to everything without an understanding of the problem.  One of the problems with data science is that models are built without sufficient human oversight of the details, for various reasons.  Making it about the demographics of the data scientists implies that there is too much oversight, and that it is biased and misguided due to the undesirable race and gender of the humans involved.

[Oexmelin]

It simply suggests that data science behaves likes most other human endeavors, where the sort of questions one asks tend to be quite influenced by background, class, gender, etc. These sorts of considerations have been shown to have an unexpected influence on supposedly value-neutral fields, like the famous example of face-recognition. It would seem even more relevant for matters such as  ethics. It has nothing to do about the nefarious intent of crypto-racists chauvinists within data science - and I don't think it was implied at all - but rather simply the sort of structural issues reinforced by uniformity of background. Data science is, if I recall, one of the least diverse fields within STEMs.

[DGuller]

It simply suggests that data science behaves likes most other human endeavors, where the sort of questions one asks tend to be quite influenced by background, class, gender, etc. These sorts of considerations have been shown to have an unexpected influence on supposedly value-neutral fields, like the famous example of face-recognition. It would seem even more relevant for matters such as  ethics. It has nothing to do about the nefarious intent of crypto-racists chauvinists within data science - and I don't think it was implied at all - but rather simply the sort of structural issues reinforced by uniformity of background. Data science is, if I recall, one of the least diverse fields within STEMs.

Can you elaborate on how lack of diversity led to problems with facial recognition models?  No general vague statements, no links to articles, can you please describe in your own words what your understanding is of the problems that lack of diversity caused?  The reason I'm asking you to not just link me to some articles is that I have read plenty of them already on this topic, and I haven't found any of them to contain any insight as to how data science is actually done.

[Oexmelin]

With that tone, I am not sure I am terribly inclined to play your little game.

[DGuller]

With that tone, I am not sure I am terribly inclined to play your little game.

Don't worry, I was well aware that you had nothing to offer even if you were inclined.  I'm glad I saved you the effort of writing four content-free paragraphs, and that I saved myself the effort of having to read them.

[Jacob]

I for one would be very interested in getting a better understanding in where the issues are in data science from an inside perspective.

A agree that vague "it's the demographics" is not a particular useful or informative answer. At the same time, I believe that unconscious bias is a real thing with real effects, and I have no reason to believe that data scientists as a group are immune to it (especially if we are positing some data scientists who are less than concerned about ethical behaviour who are also subject to pressure and performance metrics from outside the data science field).

But I'm guessing that's not going to happen right here right now.

[Oexmelin]

With that tone, I am not sure I am terribly inclined to play your little game.

Don't worry, I was well aware that you had nothing to offer even if you were inclined.  I'm glad I saved you the effort of writing four content-free paragraphs, and that I saved myself the effort of having to read them.

Go fuck yourself. I have enough self-satisfied pricks to deal with in academia without adding yet another one here. Have a wonderful day.

[DGuller]

I for one would be very interested in getting a better understanding in where the issues are in data science from an inside perspective.

A agree that vague "it's the demographics" is not a particular useful or informative answer. At the same time, I believe that unconscious bias is a real thing with real effects, and I have no reason to believe that data scientists as a group are immune to it (especially if we are positing some data scientists who are less than concerned about ethical behaviour who are also subject to pressure and performance metrics from outside the data science field).

But I'm guessing that's not going to happen right here right now.

I already gave a big part of the answer.  It doesn't matter whether data scientists have unconscious bias or not (a term which I find is used too freely, but that's a separate discussion).  The problem with the "it's the demographics" argument is that a lot of things are on auto-pilot, and the auto-pilot doesn't have a race or gender.  The reason the woman that miscarried keeps getting the ads for the babies is not because some male data scientist has no understanding of pregnancies; she's getting them because the algorithm doesn't know pregnancy from a screwdriver, and on average it works.  There is nowhere near sufficient intervention to make the algorithm more complicated and probably less effective overall by making sure that pregnancy-related suggestions are handled in a sensitive manner.

Another big problem is that things are done very quickly and very sloppily, for the most part.  There isn't enough of an engineering culture to make sure that details are taken care of with due diligence, and there aren't enough skilled data scientists around to put it in practice.  Sloppy algorithms still work most of the time, even if not to the maximum effectiveness, but they can fail rather disastrously, especially when the domain on which they're applied changes in a subtle way.