Neural nets more accurate than humans at detecting sexual orientation in images(psycnet.apa.org)
psycnet.apa.org
Neural nets more accurate than humans at detecting sexual orientation in images
http://psycnet.apa.org/record/2018-03783-002
26 comments
The fact the the authors of the paper rush for innate explanations (hormone levels in the womb...) that they have no expertise to comment on instead of much more obvious behaviors (maybe self-labelled gay and straight people upload different types of pics or take care of themselves differently, on average), is a clear violation of Occam's razor.
The "average gay face" is similar to the "average straight face" with glasses on, a better angle, and a couple of pounds less.
The "average gay face" is similar to the "average straight face" with glasses on, a better angle, and a couple of pounds less.
Exactly. The authors don't seem to have any basis for their explanations, and the title should really be "neural nets more accurate... on dating website images". Still this is significant, because these are the same images a government or business could potentially access for their own ends.
This is probaly the best article that you need to read after reading another one of those 'detecting sexual orientation with dating profile pictures/prison inmates/etc' papers:
https://medium.com/@blaisea/do-algorithms-reveal-sexual-orie...
https://medium.com/@blaisea/do-algorithms-reveal-sexual-orie...
"Given a single facial image, a classifier could correctly distinguish between gay and heterosexual men in 81% of cases, and in 71% of cases for women"
I can guess if someone of gay or straight with 98% accuracy - just always guess they are straight.
Tha ratio of gay to straight in the test data set is perhaps the most important part in determining how well the algorithm actually performs.
Are we supposed to assume the computer was shown a 50/50 split of gay and hetero people? It appears that way. But please, tell us.
I can guess if someone of gay or straight with 98% accuracy - just always guess they are straight.
Tha ratio of gay to straight in the test data set is perhaps the most important part in determining how well the algorithm actually performs.
Are we supposed to assume the computer was shown a 50/50 split of gay and hetero people? It appears that way. But please, tell us.
If the test data does not have a 50/50 split (or something around that ratio), the headline is straight up lies.
Skewed test data is the most common problem with research, reminds me of this great video by Veritasium (Is most published research wrong?): https://www.youtube.com/watch?v=42QuXLucH3Q
Skewed test data is the most common problem with research, reminds me of this great video by Veritasium (Is most published research wrong?): https://www.youtube.com/watch?v=42QuXLucH3Q
From the author's notes: "When presented with a pair of participants, one gay and one straight, the algorithm could correctly distinguish between them 91% of the time for men and 83% of the time for women."
So yes, 50-50.
The data used is from dating web sites. That's not a good sample source.
True, but they also validated it against generic facebook images; and compared the accuracy of humans on the (dating website) images to the established benchmark of human capability (which came out comparable). If there was something telling sexuality-wise from the dating pictures, then you would probably expect humans to be able to pick up on it as well.
(see author notes on this: https://docs.google.com/document/d/11oGZ1Ke3wK9E3BtOFfGfUQuu...)
(see author notes on this: https://docs.google.com/document/d/11oGZ1Ke3wK9E3BtOFfGfUQuu...)
Maybe not but still impressive that they beat humans, no?
Well, assuming the data isn't good — no, obviouly. What "beating humans" would even mean in that case? Imagine you have a dataset of dog and squirrel pictures, with "dog" and "squirrel" (or, actually, "elephant" and "shark") labels assigned quasi-randomly. You make humans and NNs to compete in what label you assigned to them. NN beats humans. Now, what the fuck that result even means?
yeah but I guess the data isn't bad in that way in that the sexual orientation of people on the site is probably labeled accurately (in fact I would expect there it's more accurate than, you don't want to lie about what you want on the dating service), the data is bad in that it is perhaps not a representative sample of the population as a whole?
What do you mean by labelled accurately? The AI only predicts how an individual self-labelled on a fairly public dating service, not how this person feels or behaves in any other way. There are so many headless torsos on Grindr (or any other gay hookup app) because many men don't want to self label as gay when they look for sex with other men. These men could very well be looking for women on Tinder.
ok, I hadn't considered that.
> is probably labeled accurately
> the data is bad in that it is perhaps not a representative sample of the population as a whole?
Yeah that's my thinking too
> the data is bad in that it is perhaps not a representative sample of the population as a whole?
Yeah that's my thinking too
that made me realize,
if you had a bad data set and let the computer train beforehand with the learning data, the human will need to "retrain" as they go (eventually realizing past stereotypes are failing in this atypical data set) and of course are going to do shit.
In a fair test both humans and computers would get to study the learning data.
if you had a bad data set and let the computer train beforehand with the learning data, the human will need to "retrain" as they go (eventually realizing past stereotypes are failing in this atypical data set) and of course are going to do shit.
In a fair test both humans and computers would get to study the learning data.
Lots of AI journalism refers to "beating humans" but that's a hopelessly vague and unhelpful comparison. No AI can ever beat a human if by "human" you mean a human equipped with whatever tools the human may choose to use, because those tools would include the AI against which you would have it compete. So the comparison only makes sense if you restrict the human in some way. Depending on how you restrict the human, beating the human may be very easy or very hard. For example, if you insist that the decision be made within 100 ms then beating the human is very easy.
I think the usual way in which these experiments fiddle things so that the AI wins is by giving the AI lots of training data but giving the humans no training whatsoever, which may seem grossly unfair, but apparently that's acceptable in AI research/journalism.
I think the usual way in which these experiments fiddle things so that the AI wins is by giving the AI lots of training data but giving the humans no training whatsoever, which may seem grossly unfair, but apparently that's acceptable in AI research/journalism.
Me: Horses can run faster than people!
You: "Faster than people"?! That's a hopelessly vague and unhelpful comparison. No horse can ever go faster than a human if by "human" you mean a human equipped with whatever tools the human may choose to use, because those tools would include the horse against which you would have it compete. So the comparison only makes sense if you restrict the human in some way.
Me: Um..
:-) The other stuff you say makes perfect sense, but that first bit is more pedantic than fair, I think.
You: "Faster than people"?! That's a hopelessly vague and unhelpful comparison. No horse can ever go faster than a human if by "human" you mean a human equipped with whatever tools the human may choose to use, because those tools would include the horse against which you would have it compete. So the comparison only makes sense if you restrict the human in some way.
Me: Um..
:-) The other stuff you say makes perfect sense, but that first bit is more pedantic than fair, I think.
You make a good point. There are many situations in which it's fairly obvious how the comparison ought to be made: the human should not be allowed to use a horse in competing with the horse in "running", and should not be allowed to use a computer in competing with the computer in "guessing sexuality". However, if the competition is in "predicting share price movements" then surely it would be strange not to allow some kind of computational assistance for the human, so in those cases, at least, it gets a bit hard to decide where to draw the line.
A bit unrelated but there are horse vs people marathons. and many times people win. that is because horses cannot cool themselves like people can
Yes, see "persistence hunting". It turns out that although humans may be thrashed by a poodle over short distances they are really rather good at extreme-long-distance races.
Some additional context and opinion on this:
https://www.washingtonpost.com/opinions/ai-gaydar-could-comp...
They will be using it to detect criminals next.
But why does it matter? What could it possibly be used for? I can't think of a use for this which wouldn't be crossing a line / inappropriate / dangerous.
Homosexual relations are illegal in ~70 countries, and punishable by death in ~10 of those, such as Iran, Saudi Arabia, Brunei, Yemen, etc. Governments in these countries could run this software in their databases, rooting out and prosecuting or even executing those identified. So I would say this kind of software merits discussion.
The governments in these countries are evil, not stupid. The false positive rate is too high for this to be useful at anything but a really coarse, maybe check this guy out level.
All of the countries you name bar Yemen are easily capable of rooting out and destroying their gay communities if they really wanted to. They don’t. What they want is to make sure everyone involved is very discreet. Hypocrisy is the homage vice pats to virtue etc.
All of the countries you name bar Yemen are easily capable of rooting out and destroying their gay communities if they really wanted to. They don’t. What they want is to make sure everyone involved is very discreet. Hypocrisy is the homage vice pats to virtue etc.
- https://news.ycombinator.com/item?id=15198997
- https://news.ycombinator.com/item?id=15197287
The author of this paper has created a document that addresses:
(1) Summary of the findings
(2) You must be wrong – this is pseudoscience! (common criticism of this paper)
(3) Our response to an irresponsible press release by GLAAD and HRC; or, a much better response by LGBTQ Nation
You can find that here: https://docs.google.com/document/d/11oGZ1Ke3wK9E3BtOFfGfUQuu...