Row over AI that 'identifies gay faces'
bbc.co.uk
bbc.co.uk
Edit: Can anyone downvoting me please stand up and tell me why?
Is the assertion here that lies from gay people should be ignored?
It's pretty normal for charities to encourage donations, so I'm not sure what your problem is with that. You seem to be hinting at the idea that they have deliberately released an inaccurate press release to encourage donations, but there's no reason for them to bother doing that when there are many uncontroversial instances of homophobia that they could refer to.
On a more specific note, claims like "The study assumed there were only two sexual orientations -- gay and straight -- and does not address bisexual individuals." (without mention that this is addressed in the paper and that better labels would likely improve results) and repeated hand-waving about hairstyles (without mention of the sections of the paper devoted to feature analysis which point to structural facial features giving the strongest signal to the model) do not present methodological flaws with the study and/or are misleading. Prof. Kosinski and Yilun Wang have issued a very polite rebuttal (here)[https://docs.google.com/document/d/1UuEcSNFMduIaf0cOWdWbOV3N...] which details these and many other weaknesses with the press release. I don't think it's unreasonable to expect someone at GLAAD or HRC to understand what peer review means, and in fact, I'm sure they do.
The last thing an issue like climate change needs is a falsifiable hypothesis which is why the IPCC never puts one forward.
If we want people to stop driving cars, take shorter showers and do other progressive things than the last thing we should do is propose something that could be falsified. It would be as devastating to climate change as the catholic church putting forth a testable hypothesis for the existence of God.
In fact, the original paper is deeply confused about sexual orientation:
>Their sexual orientation was established based on the gender of the partners that they were looking for (according to their profiles).
> [...]
> Another issue pertains to the quality of the ground truth: it is possible that some of the users categorized as heterosexual were, in fact, gay or bisexual (or vice versa). However, we believe that people voluntarily seeking partners on the dating website have little incentive to misrepresent their sexual orientation.
Obviously, bisexual people who say that they are looking for a partner of one or other gender are not "misrepresenting" their sexual orientation!
As with all organizations the purpose of organizations are to perpetuate themselves.
On peer review: 1) The uploaded version of the preprint that was linked is currently version 7. Version 1 did not contain the "this is a peer reviewed article" boiler-plate at the top. It's very possible that GLAAD/HRC saw the original manuscript without this notice. 2) Given that all 7 versions have been posted in the past 9 days, one might reasonably conclude that this is still a work-in-progress and peer review is still currently on-going, (the boilerplate being added in anticipation of the review being completed. Again, the paper is very clearly still being worked on.)
On misinformation: Yes I will give you that some of the criticisms might seem over-the-top to the HN crowd, but I think they are understandable given what GLAAD's / HRC's focus is, plus the fact that they aren't addressing "trained scientists." That being said, I don't think their criticisms are completely baseless. I think anyone with technical training, particularly in data analysis / machine learning, is familiar with the phrase: "Garbage In - Garbage Out." And I think it is absolutely correct to point out that maybe if your training data is dating site photos, predominantly of white people in a narrow age range, that you may not have exactly proven your claims that you have "advance[d] our understanding of the origins of sexual orientation and the limits of human perception." That maybe you've just learned some interesting features of current trends in dating site photos.
The tragic thing is that both sides (GLAAD & HRC vs. the researchers) claim to care about the privacy and safety of individuals. And kudos to them for thinking of the social implications of this kind of research! (We all know how good phrenology turned out) Because of these implications, I absolutely think it's right for us to question the conclusions of this research (that you can detect sexual orientation from facial morphology, and that this is evidence of the prenatal hormone theory of sexual orientation), given their actual methods (we threw a bunch of dating site photos at a neural-net classifier, and compared to self-reported sexual orientation)
All that said, I do think that the paper is likely to be more nuanced than the conclusions reported / the abstract make it out to be (have not finished reading the paper yet). But unfortunately, because of the way science is reported, where the headline is going to be some bombastic "AI can tell if you are gay or straight from just a picture!", there is probably some pressure to voice their criticisms just as loudly.
Again, I hate to keep parroting what another HN said, but I hate to see misinformation repeated:
>The paper states that they used an existing net, VGG-Face, to exact facial features: https://osf.io/fk3xr/ (page 13) >VGG-Face aims at representing a given face as a vector of scores that are as unaffected as possible by facial expression, background, lighting, head orientation, image properties such as brightness or contrast, and other factors that can vary across different images of the same person. All the heavy lifting for this paper was done by logistic regression on the facial features reported by VGG-Face, they didn't train a DNN specifically for identifying sexuality. The algo wouldn't have seen clothing at all.
Basically, it's not as simple as "All the gay participants had rainbow flags in the background of their photo so let's ignore this study"
What I am saying is that the authors may have been a bit far-reaching in their conclusions, and that "trends in dating site photos" doesn't just mean "is there a rainbow flag present?", but also things like how photos are taken w.r.t angles, and lighting, and retouched, makeup, expressions, etc. (and yes VGG-Face is supposed to be resilient to some of these things, they also used Face++ for studies having to do with morphology, such as using facial contour features. But saying they are resilient to these things and quantifying that are two very different things).
To get to even a more fundamental question - Does the population of gay and straight individuals on a dating website necessarily reflect the gay and/or straight populations as a whole?
And once the authors start talking about the femininity or masculinity of faces, and whether certain features are gender atypical or not, they are entering very murky and ill-defined territory, which the authors even acknowledged when relating previous results on this front.
To be very clear about it: I am not saying that the study should be ignored completely, or that there were no interesting findings. Quite the contrary! What I am saying is that the conclusions being reached ("Our results provide strong support for the PHT, which argues that same gender sexual orientation stems from the underexposure of male fetuses and overexposure of female fetuses to prenatal androgen responsible for the sexual differentiation of faces, preferences, and behavior") are quite a bit of a leap from what was found in the research given the input data and methods, and I think should have been more accurately reported as: "interesting differences in social media photographs of gay vs straight individuals.", with a little less speculation by the authors on what might be causing these differences.
Funny thing though is R=0.19 being used as evidence. First of all, effect size is what matters. Second, this is near zero and essentially worthless. Third, they didn't really check how their methods work or know why they work, just guess.
For fun, let's see how easily you can make someone gay with evolutionary approach used in the 2015 fooling face detectors paper. The algorithms are in opencv_contrib.
I bet it is essentially worthless.
The inputs are self-submitted photos to a dating website. Can we be certain that how people pose, smile, look, etc, isn't influenced by their intention- to find a mate of a specific gender? What if those same people were each asked to submit a second photo when they're told to 'act gay/straight' in order to fool the algorithm?
It could be that this ML algorithm is picking up on very different cues than intended. But I wouldn't discourage the researchers over it- try it again with a better data set, one built for this purpose, and discover just what the algorithm is picking up.
The paper states that they used an existing net, VGG-Face, to exact facial features: https://osf.io/fk3xr/ (page 13) >VGG-Face aims at representing a given face as a vector of scores that are as unaffected as possible by facial expression, background, lighting, head orientation, image properties such as brightness or contrast, and other factors that can vary across different images of the same person. All the heavy lifting for this paper was done by logistic regression on the facial features reported by VGG-Face, they didn't train a DNN specifically for identifying sexuality. The algo wouldn't have seen clothing at all.
Basically, it's not as simple as "All the gay participants had rainbow flags in the background of their photo so let's ignore this study"
Well, you've made me laugh and made an excellent point.
I suppose I should read the paper more in-depth.
Is it that doubtful that a ML algorithm can reach/exceed human-levels on this "gaydar" problem?
I get that this study is problematic, but there is no reason to automatically assume it's not possible. What evidence do we have that there is no correlation between ML consumable physical characteristics and our sexuality?
"Bad actors" using AI to influence debate, or writing "fake news" then promoting it with social media ads, etc?
Yes, and it is actually the bigger problem. Facial recognition and license plate scanners are already producing a "papers, please" situation.
The gay community, however, is more activist and more likely to fight back. Which is why this is getting such traction.
"Additionally, given that companies and governments are increasingly usingcomputer vision algorithms to detect people’s intimate traits, our findings expose a threat to the privacy and safety of gay men and women." (page 2)
The setup and modeling may be perfectly sound, but the conclusions are a big stretch: they didn't control for intentional differences in presentation (the data was dating site photos), nor did they appropriately test for significance of the supposedly subtle facial differences picked up by the model (gay men having more feminine faces, gay women more masculine ones).
The paper states that they used an existing net, VGG-Face, to exact facial features: https://osf.io/fk3xr/ (page 13) >VGG-Face aims at representing a given face as a vector of scores that are as unaffected as possible by facial expression, background, lighting, head orientation, image properties such as brightness or contrast, and other factors that can vary across different images of the same person. All the heavy lifting for this paper by was done logistic regression on the facial features reported by VGG-Face, they didn't train a DNN specifically for identifying sexuality. The algo wouldn't have seen clothing at all.
It's pretty disturbing how quick people are to dismiss papers that make them uncomfortable, latching onto the first excuse to dismiss it without even confirming if the excuse is valid or not. Clear Cognitive Dissonance in action.
The first is that claiming this methodology outperforms humans is unfair; there's no expectation that Turkers are particularly good at identifying sexuality from photographs, they haven't been able to train in a remotely analogous way, and there's no comparison to "experts" in orientation recognition (if those exist).
The second is that the authors of this study use the differences extracted by VGG-Face to claim support for PHT, despite not applying an appropriate statistical test to the differences in the features between groups. This is the bigger scientific misstep in my mind, the willingness to make a strong, apparently controversial claim (I'm not too familiar with PHT or its history) without properly validating it.
We know that hormonal derivatives like steroids can influence appearance, and also that ambient secretions offer signaling at a distance among humans, in circumstances not limited to the synchronization of menstrual cycles, at a minimum.
Take these factors together, and exclusive close proximity to gender concentrations based on preference likely results in mild changes in appearance that automated analysis can determine as a non-zero signal, and hazard an informed guess with. People are not wrong in claiming that this is a form of stereotyping, and that sometimes the stereotype will be incorrect.
People may not like it, but just because you can't see it, doesn't mean it's not there.
Sexuality is not a boolean value, it's not a yes or no, it's not a this or that, even for people who are attracted to females they are not attracted to every female, and people who are attracted to males are not attracted to every male. It's such a silly thing to try and hone in on, because it removes the primary deciding component of human sexuality from the study: personal, intimate, real human preference. It's like attempting to discover one's favorite flavor of bagel shmear based on high school portraits.
You may find some surprising results, realize neural nets are trainable phenomena, and then run around giving everyone plastic surgery because more pronounced cheek bones tend to prefer Strawberry Shmear, and if peoples' cheek bones were adjusted their flavor preference would change.
No, it doesn't. It means sexual orientation and hormonal/chemical composition might be related, and hormonal/chemical composition might affect facial composition.
Everything you wrote afterwards is just an emotional outburst to an uncomfortable results from a study you obviously didn't bother reading.
Yes, I am emotionally bothered by this whole enterprise, because sexual orientation is not something you can understand through portraits, and this study makes the claim that 7/10 times a computer can do it. That's not the case.
Let's say there's a 50% chance that hormonal/chemical composition and sexual orientation are related, and let's say there's a 50% chance that hormonal composition affects facial features. That means that for the both of them to work out we have a 25% chance they are correlated. That's a pretty generous estimate, given that there is no data supporting the second half of that equation where "hormonal composition affects facial composition" as far as I'm aware.
The best we are doing with this dataset is inferring hormonal composition from what someone put down as their sexual preference. Or am I missing a large piece of the puzzle?
Why couldn't physical features change?
I believe hormone changes can affect appearances, so it doesn't seem beyond the realm of possibility.
> Sexuality is not a boolean value, it's not a yes or no, it's not a this or that
Did it occur to you that possibly the correlation with facial characteristics could be a matter of degree?
Consider the correlation between facial characteristics and racial background. I'm half European, half Asian. There's going to be a stronger correlation between people with near 100% European ancestry and "European facial characteristics" than people with 50% European ancestry, than people with 25% European ancestry, etc
You don't even have to take issue with the modeling or any other part of the methodology to debunk this study; it's the interpretation that is plainly ridiculous. As you said, these are pictures from a dating site; awfully likely that people choose to present themselves differently based on who they want to attract. Also no appropriate statistical test applied to differences in faces picked up by the model. See this for more: http://callingbullshit.org/case_studies/case_study_ml_sexual...
What's the big deal? We've already got people thinking their toddlers are transgendered because they picked up a doll one time...
I am certain that you could introduce a lot of errors by adding a bunch of metrosexual but straight individuals into the mix.
>VGG-Face aims at representing a given face as a vector of scores that are as unaffected as possible by facial expression, background, lighting, head orientation, image properties such as brightness or contrast, and other factors that can vary across different images of the same person.
All the heavy lifting for this paper was done by logistic regression on the facial features reported by VGG-Face, they didn't train a DNN specifically for identifying sexuality from image pixel values. The algo wouldn't have seen clothing at all.
I still think that while the results are accurate(ish), they reveal fashion trends and not genetic changes. Can in point, the composite average of homosexual male has glasses. Facial features of homosexual are often gender atypical but that's trough a lot of effort to conform with fashion trends (eyebrow reshaping, smoother skin, "kinkier" smile, etc.). This is especially true on dating sites. The whole nature vs nurture thing.
While I agree that hormone levels can impact facial features, I have yet to see a strong study proving that hormones levels can be directly linked to sexual orientation or gender identity.
"But face recognition is a sensitive/politically-charged topic, I know several grad students (including me) who inspite of having ready-to-deploy scalable software and datasets (~10M instagram images) stay away from doing this type of a demo because reputational risks are enormous. Consider the controversy around geofeedia etc. The last thing you want as a PhD student is press interpreting your research incorrectly and blaming you for causing widespread harm. It happened to a student/professor in my department and even then the infamous study in question was in collaboration with the social network."
I think the reason this topic is so hot is not because of the study, but the "What if other people use this as a justification for X?" fear.
For example, if it turns out that there are certain physical traits that predict sexuality, does this imply a physiological component (rather than a purely psychological one)? If there is a physiological component, I think some people worry this can be used to "pathologize" alternative sexualities.
However, what if all the endocrine disrupters we're exposed to daily [1] that act as estrogens play some role in this? What if we make the topic so unspeakable that we fail to recognize environmental contamination causes both physical differences [2] and sexual differentiation differences [3]?
This is part of the reason I think we can't reject information that may allow some socially-backward conclusion. Because we don't know all the other good things (i.e. medical) that information may be necessary for.
1. https://en.wikipedia.org/wiki/Bisphenol_A#Uses 2. https://en.wikipedia.org/wiki/Anogenital_distance#In_humans 3. https://en.wikipedia.org/wiki/Diethylstilbestrol#DES_sons
However, it's both interesting and dangerous, as the GLAAD press release states, because of the repercussions to gays in countries where it's not nearly as accepted. It probably won't stop Russia and other very homophobic countries from conducting the same experiments using their own datasets.
In addition, very deep research including gene sequencing might inadvertently turn up the "gay" gene, which would then lead to it's own Pandora's box of issues.
This might be one of those situations where we might not want to study too hard because of the repercussions associated with it.
(I know you didn't claim we did, just pointing out that your examples don't seem to support your argument.)