2 ms·
I can't read the WP article, but I assume it reports on the Deep Gaydar paper, from Stanford (https://osf.io/zn79k/ https://osf.io/zn79k/). That paper is sever
by YeGoblynQueenne 8y ago
I can't read the WP article, but I assume it reports on the Deep Gaydar paper,
from Stanford (https://osf.io/zn79k/ https://osf.io/zn79k/).
That paper is severely flawed, first of all in the way they trained their
classifier to recognise peoples' sexuality from images of their faces.
What they did is they collected ~35k pictures of gay and straight men and
women, all users of a dating site in the US. Half of them were gay and half
were straight.
Right off the bat, that breaks the fundamental assumption in machine learning,
that of distributional consistency. In short, machine learning works under the
assumption that the distribution from which a training sample is drawn is the
same as that of unseen data- i.e., of the real world. If the sample
distribution is not the same as that of the unseen data, a model cannot be
guaranteed to predict anything about unseen data.
But, in the Deep Gaydar paper's sample, the distribution of gay and straight
men and women was uniform: 50% straight and 50% gay people, while in the real world the prevalence of homosexual preference is estimated to be around
10%- the Deep Gaydar paper itself cites a 7% prevalence.
So the classifier trained to identify peoples' sexuality from images of their
faces, in the paper, is completely incapable of identifying anything about the
sexualities of people outside its training sample- which is the majority of
people on earth. Which invalidates their claim that it's possible to tell
someones' sexuality from an image of their face. At best, what they can claim
is that they trained a classifier to identify the sexuality of people in their
training sample from images of their faces. But nothing more.
And no, that does not suggest that "physiognomy is not pseudoscience". Quite
the contrary- sneaky data manipulation like that is the hallmark of crackpots
and charlatans.