Read on Twitter
Went down something of a "Race in AI" rabbit hole last night, & am reading a paper by Gebru (of fired-by-Google & led-a-woke-mob-against-Yann-LeCun fame), & if you know anything about ML & have spent any time at all in the gender self-ID wars you'll immediately spot the landmine:
You can't insist that machines be trained on datasets that include representation of "male" & "female" faces, & ALSO insist that these gender categories are purely a product of self-ID w/ no relationship to outward presentation. In other words, if "don't assume my gender" is...
going to be A Thing, then there is zero rationale for training machines to assume someone's gender! It's physically impossible if gender is purely an inner state that may or may not affect outward presentation, & socially a non-starter if "assuming" gender is verboten.
So arguments about gender bias in AI image recognition are fundamentally moot in a world of gender self-id. The whole debate is a no-op. What would it even mean to train a model on two different pictures of Elliot (formerly Ellen) taken the day before & the day after transition?
What category of error is it if an ML model classifies this picture as "female"? If you search for pictures of women actors, should this come up in the results or not?
I think the "solve" here (to use startup speak) is the elimination of these gender categories entirely for a crude set of biologically descriptive ones -- exactly what's going on with this "vulva owner" & "menstruator" & "penis owner" talk in media outlets, now.
If your brain exploded because Healthline is now refusing to speak of "women" & instead uses "vulva owners", wait until Google replies to your attempt at an image search for "women" with "Did you mean vulva owners?"
I guess what it boils down to, is that unless an ML can infer extremely subjective inner mental beliefs from a photo or video of you, then (self-IDed) gender categories are not possible to pattern-match. And if an ML /can/ infer such, then we have an even bigger set of problems.
