AI that can determine a person’s sexuality from photos

Eriol

Eriol

I really am not meaning to exaggerate in that title, but that's what the article says: AI that can determine a person’s sexuality from photos shows the dark side of the data age
When presented with multiple pictures of a pair of faces, one gay and one straight, the algorithm could determine which was which 91 percent of the time with men and 83 percent of the time with women. People provided the same images were correct 61 and 54 percent of the time, respectively — not much better than flipping a coin.
Click to expand...
91% and 83% is pretty good. Not perfect of course, but... frighteningly accurate. But I need this paragraph explained:
This accuracy, it must be noted, is only in the system’s ideal situation of choosing between two people, one of whom is known to be gay. When the system evaluated a group of 1,000 faces, only 7 percent of which belonged to gay people (in order to be more representative of the actual proportion of the population), it did relatively poorly. Only its top 10 showed a 90 percent hit rate.
Click to expand...
Top 10% what? I don't get that paragraph.

And the photos were from a DB where the people classified THEMSELVES so it's not researchers cherry-picking what "they thought" sexuality looked like.

It's in this forum because there's no way this can't turn political in SOME way or another.
 
blotsfan

blotsfan

The fact that it uses self-identification means that it only knows gay people that are out. This seems like it'd make sense for there to be a difference as gay culture definitely has different ways of expressing itself in appearance (obviously there are exceptions).

I wonder how it'd do with gay people from foreign countries where I'd assume the culture is different.
 
PatrThom

PatrThom

I remember the case of a neural network that was trained to recognize whether or not there were military tanks in a picture. It was trained until it achieved 100% success, but then when it was tried on "real" photos, it failed miserably. Turned out the network had trained itself based on the fact that all the sample tank pictures were taken on cloudy days.
Any time I hear about "...can identify <blah> based on pictures," I remember this incident, and have to question exactly what the AI has decided are the "tells."

--Patrick
 
Eriol

Eriol

PatrThom said:
I remember the case of a neural network that was trained to recognize whether or not there were military tanks in a picture. It was trained until it achieved 100% success, but then when it was tried on "real" photos, it failed miserably. Turned out the network had trained itself based on the fact that all the sample tank pictures were taken on cloudy days.
Any time I hear about "...can identify <blah> based on pictures," I remember this incident, and have to question exactly what the AI has decided are the "tells."
Click to expand...
I was given that exact example in a neural networking class.... 13 years ago. Training sets are always a thing, but given that it's from a public database in this case, that kind of thing gets less likely.
GasBandit said:
Gaydar is now a thing.
Click to expand...
I ALMOST used that word in the title, but decided against it.
 
@Li3n

@Li3n

Mah, sounds like they picked their original subjects based on how they looked compared to the stereotypes. So of course IRL, where there's nothing that actually makes you like a certain sex that also makes you dress a certain way, it's not so reliable any more.
 
GasBandit

GasBandit

Staff member
I get the feeling that whatever it measures for "gayness" is only useful in a comparative situation, and not in a standalone evaluation. It said, the machine did well when presented with two pictures, and asked "which of these two people are gay" but not when being presented with one picture and being asked "is this person gay." It might just be comparing something like "skin smoothness" or something and guessing whoever has the smoother skin is gay, or some other such measure that only works when comparing two pictures. With one picture alone, it has to think "is this person's skin smooth enough to be gay?" and without comparison references, that's a much more difficult call, even if the criteria is valid, which it might actually not be.
 
Dave

Dave

Staff member
Like are these headshots or pictures of porn? Because if it's pictures of porn I'm like 85 - 90% effective myself.
 
You must log in or register to reply here.
Top