TheContact
Member
This is kinda scary I can see this being used to discriminate
It's not just a landmark classifier. They also used VGG-Face, which appears to be a CNN feature extractor, reduced features with SVD and then ran it through logistic regression. This isn't all that different then just running a CNN directly on the image.
And Figure 3 shows what I was talking about regarding the masking thing. It shows parts of the image that are important to the CNN feature extractor.
This is kinda scary I can see this being used to discriminate
Furthermore, in an even more rigorously controlled series of experiments published in the Journal of Personality and Social Psychology, Rule and his colleagues replicated their discovery that people are able to accurately guess male sexual orientation. This time, the researchers demonstrated that perceivers were able to do this even when they were shown only individual features of the target's face. For example, when shown only the eye region ("without brows and cropped to the outer canthi so that not even "crow's-feet" were visible"), perceivers were amazingly still able to accurately identify a man as being gay. The same happened when shown the mouth region alone.
I see. Thanks for reading the actual paper. It's actually pretty impressive then, given they presented it with equal numbers of heterosexual and homosexual people.They took equal numbers of gay and heterosexual people (3947 heterosexual men, 3947 homosexual men, 3441 heterosexual women, 3441 homosexual women). They were then presented to a neural network in pairs of one heterosexual and one homosexual to train in and get some accuracy statistics. Because the program gave confidence bounds, there is no single convenient precision/recall statistic.
However, here is the relevant part from the study itself:
I understood VGG Face to be the landmark extractor that was taking measurements of control points. It didn't sound like it was pulling out arbitrary features but actually measuring things and was trained independently of the rest of the network for that task. That would make sense because if you want to say that gay people have different nose bridge sizes then it should be obvious by measuring and plotting which can then be classified with a simple logistic classifier. Otherwise I have no idea how you draw conclusions if it's picking all the features for you, you know it's around the nose, but does that mean they have a piercing or lack a mustache or something?
What? That's not very impressive, if you just guess everyone is straight, you'd be correct 95% of the time.
They took equal numbers of gay and heterosexual people (3947 heterosexual men, 3947 homosexual men, 3441 heterosexual women, 3441 homosexual women). They were then presented to a neural network in pairs of one heterosexual and one homosexual to train in and get some accuracy statistics. Because the program gave confidence bounds, there is no single convenient precision/recall statistic.
However, here is the relevant part from the study itself:
This is kinda scary I can see this being used to discriminate
Here:
In fact, I would read the last two pages of this thread. They would probably answer any questions about the methodology of this study that may pop up.
But it's still relevant -- creating a contrived situation where the base rate is lower doesn't mean very much as to the real-world applicability of the technique.
You know what. This literally proves that being "gay" isn't a "choice". The physical body shows "traits" we may not even notice...and I find that pretty awesome.
Rather than this PROVING that to horrible human beings they would use it against people I suppose....derp.
Like it said, it's not 100%...but I find the things the AI picks up on and uses facinating (and pretty obvious).
90+% is darn impressive.
In other words I would guess this isn't a "gay detection machine" you can point at any photo to determine secrets, something that would make politicians or religious people nervous. They're likely not going to look gay to the AI, and will say "nope I'm not gay, see, it works." The people who look gay to the AI are the kind of people who would tell the scientists and researchers "yes, I am gay."
They actually address this in the author's notes:
Like they say, it's obviously very hard to test on closeted people with a statistically significant sample size, but it's not unreasonable to assume it would work.
So erm how is this useful to humanity?
Fascinating. But what's the purpose of this research? Just cos?
Does it have to be?
It is a strong evidence that being gay/straight may not be a thing that you just learn and that people don't born neutral/empty.It basically shows that (at least in part of course) there are biological mechanisms pushing people to certain behaviours. This is basically what the likes of steven pinker were saying for more than a decade now.
And your point is?What an odd bump after 2 weeks of this not having been posted in.
The post this is in reply to is 16 days old and had numerous responses.