I'm not sure where GP is going with this, but for example the Habsburger lower lip [1] is a easily identifiable feature from inbreeding in Western European nobles. If you can identify it it gives you a good clue about the socioeconomic class of the person, which gives you a good clue about their political orientation.
I'm not sure this is enough. If you classify all old white men as Republican you can have 69% accuracy. I'd even argue the more such factors you control the less it means something.
I think you misunderstand what "controlled for" means in this case. The way the controlled version of the test works is that two pictures are presented at the same time. Both subjects will be of the same gender, the same ethnicity, and (approximately) the same age. The goal of the test is to choose which of the subjects pictured is conservative, and which is not. There is no way to obtain greater than 50% accuracy by choosing "old white men", because both pictures will be of "old white men" (or "young black women", or whatever). Something else in the photo is being used to obtain the boost in accuracy: perhaps facial hair, perhaps obesity, perhaps apparent youthfulness, perhaps a guess at sub-ethnicity, perhaps pose---but it has to be something other than "old white man".
I'd posit that it's some combination of genetic controls on things like openness, conscientiousness, threat processing, and the fact that people have children who look like them and on average live in groups who look somewhat similar.
Can you explain to me how that’s a bad thing?
Differences among people do not make them more United and their bonds stronger.
Literally the opposite is true.
That depends a bit. Differences can complement each other.
Some people might be better at some things, while others are better at other things. (thus combining strengths)
Some people might actually be really bad at some other things yet again, while others could happen to be very good at them (thus covering each others weaknesses)
When you have people with different backgrounds and training all working together you can do things together that you wouldn't otherwise be able to do separately.
For sure. I was just stating why echo chambers are bad. In the context of a dating app, I think dating right now suffers from a totally diffent and unrelated effect where the pool of potential partners has become so big that people have unrealistic expectations in what they want in a partner. So in that context I actually think anything that narrows the list of potential partners is a good thing.
But still, it's yet another facet of our lives where we deal with disagreements by putting them out of sight rather than communicating and understanding.
It's a terrible thing that politics influence this at all. A few months ago I was discussing this with my SO and we came to the conclusion that ~5 years ago neither of us would care about each other's political stances. Today, politics have infiltrated literally every single facet of life and became a mania of sorts for many even non-political types. Ignoring it is difficult to say the least. With that said, political differences, in my humble opinion, is a great thing as long as they're not radical and within normal realms. But, in a polarized climate we live in now radical political views are all the rage and it'd be very hard to dismiss in a relationship.
Former Kaiser of Austria-Hungary, Franz Joseph, had a personal motto of "Viribus Unitis". That means "With United Forces". It was also used by A-H military.
In practice, the multicultural empire was greatly weakened by incessant nationalist bickering.
Once you have to declare that X is your strength or something similar, it most likely isn't.
I would argue that Diversity is necessary, but not sufficient.
("E pluribus unum" is -of course- the motto of the United States of America. Not the least of nations!)
"If we all reacted the same way, we'd be predictable, and there's always more than one way to view a situation. What's true for the group is also true for the individual. It's simple: Overspecialize, and you breed in weakness. It's slow death. "
Well, it's an app for dating, not an app for finding "change my mind" political debate; so being an "echo chamber" is kind of what's desirable for that app.
We know incest increases the likelihood of genetic disease. I wonder if a dating app could make a "faces too similar" filter to reduce chances of genetic disease in offspring based on just faces.
Understanding why it is predictable and using that to either exclude ossified groups from targeted campaigns to save resources or focusing on them more tightly if they’re swing.
May be the world polarising into two kind of extreme polarisization: one with a sense of humor and one without it. And it has a noticeable effect on the face. I'd love to see few samples of faces from each class.
I think it's yet another situation where everyone is escalating hard from "hey, there are some stereotypes in this particular media that are problematic when presented uncritically".
That initial idea somehow turns into, "The left hates Dr. Seuss" and wild slippery slopes about book burning.
That isn't true. There was a scholarly paper that came out in 2019 observing that a few of Dr Seuss's many children's books included racial stereotypes that are dated to the point of being offensive, and that Geisel/Seuss had also produced a lot of adult cartoons that were overtly racist, a complicating factor for educators who used his books to to educate children about discrimination.
Recently the publisher announced it was going to stop selling those - I think it was 6 books out of 107 in the catalog. There was no pressure campaign or wave of outrage driving it.
Perhaps you should choose better news sources, as the ones you are using seem to be serving you poorly.
Ebay is banning accounts that sell Dr. Seuss books.
Fact.
The outrage over racial overtones came from a left aligned organization.
Fact.
It was carried on and cheered on by left leaning people.
Fact.
People on the right don’t use words like “problematic” and “racialized undertones”
That’s on you guys.
This is from The horses mouth:
“ Six popular Dr. Seuss books — including And to Think That I Saw It on Mulberry Street and If I Ran the Zoo — “will stop being published because of racist and insensitive imagery, the business that preserves and protects the author’s legacy said Tuesday,” The Associated Press reported.
“These books portray people in ways that are hurtful and wrong,” Dr. Seuss Enterprises told The Associated Press in a statement marking the late author and illustrator’s birthday.”
You’re right about me saying fact not making it a fact. Fact. (Oh god I don’t think I can stop now...)
However, I just shared a bunch of open source facts with you. All sourced from left leaning sources as to be somewhat unbiased (even though my bias on this subject is probably pretty heavy)
This isn’t about winning an online argument, it’s about a call for reason when things have become so silly that we are a step away from the Grinch being canceled because his big nose looks Jewish and that’s Nazism or something.
Not one single neo Nazi has ever been uplifted and supported in their beliefs by classic children’s books.
I thought that adding factual and documentary context in a polite way might elevate the conversation, but it looks like you prefer being rude and angry.
Both the far left and far right have no sense of humor because both have lost all sense of reasonableness.
On the right you have Trump cultists who loyally cling to every word he utters, regardless of it is true or not, spreading lies and misinformation on social media about voter fraud, vaccines, etc. You also have covid denial on the right where "the constitution" is used as a bludgeon to resist any reasonable public health policy like mask usage.
On the left you have the wokeness mobs that go around trying to cancel/censor/destroy every historical figure/book/statue who didn't live a perfect life and destroying careers of anyone who says the wrong word or makes a bad analogy. You also have covid zealotry on the left where "covid deaths prevented" is prioritized above all else and we can't re-open schools, re-open businesses, visit grandparents, or otherwise get back to normal life (even after mass vaccination) until 100% of all remaining unknowns are known (even if it takes another 2 years of masking, zooming, and hermiting).
This made me think a bit - I imagine most of the polarization these days comes from a vocal minority of either side, and that 95%+ of people are much more relaxed about politics (and still have a sense of humour) than we imagine from the media.
So it would be interesting to see if hyperpartisans could be separated from moderates by their appearance (just to be clear, I dont believe in phrenology, I imagine its information leaking from other aspects of presentation and background).
With respect to humour, the modern lack of humor seems to focus on not offending people, which I admit I dont understand but I could guess is predicated on the feeling that laughter is somehow equated to divisive mocking, as opposed to fun. I bring this up only because it reminds me a lot of Umberto Eco's "The Name of The Rose" where the 13th century religious leaders were arguing that laughter was inappropriate, and Jesus never laughed, somehow rooted in the belief that finding humor in things admitted the possibility of laughing at aspects of religion and therefore not taking them seriously.
> Accuracy was similar across countries (the U.S., Canada, and the UK)
This sounds more like people who are from minority groups (or who look like they are, to a computer vision algorithm) are more likely to agree with left-leaning policies, which probably has more to do with the policies in those countries than it does with any sort of genetic features. I feel like this might not work as well in non-Western countries, for example.
So your race accounts for less than 3% despite 89% of all black people voting left, that seams strange.
I did not do the full math but just the ballpark numbers I entered in to my calculator says that it should account for about 11%.
* 89% according to numbers from nbc
Edit:
So say I assume that every African American person I see votes left.
13% of the population is African American 89% of them actually vote left.
13% * 89% = 11%
Then I simply guess on all non African American a 50/50 shot.
I should then be right
50% + 11% = 61% of the time.
The study looked at political orientation (liberal versus conservative or left versus right), not political affiliation (Democrat versus republican). 89% of African Americans vote Democrat for complex historical and sociological reasons. But they have a large diversity of political opinions: https://press.princeton.edu/ideas/the-roots-of-black-politic.... About 30% of Black people today identify as conservative, versus 10% in 1970. But Democratic Party affiliation has been in the 90% range throughout that whole period.
I only skimmed the paper, so I'm not claiming to know much about it, but one thing to keep in mind here is that a fair coin has a 50% accuracy using the same terminology as the headline. I'm not saying 72% is not an interesting achievement, its just that "you can do about 50% better than random chance" describes my gut feeling about how much you could actually see in someones face.
There were tons of national polls done for Trump's approval that included all adults (instead of likely or registered voters). Trump fared noticeably worse in the polls of all adults throughout his presidency.
I wonder if that 55% is from mturk or other survey sites that can be somewhat questionable in terms of quality with how much people are paying attention versus maximizing their hourly survey earnings.
They do note the random chance bit, and they also note that it's better than humans could judge on their own and even, surprisingly, better than judged by a personality questionnaire.
> Political orientation was correctly classified in 72% of liberal–conservative face pairs, remarkably better than chance (50%), human accuracy (55%), or one afforded by a 100-item personality questionnaire (66%).
Were the humans experts or just random people though?
The real question is whether the tool can beat a lookup table of age, race, and gender probabilities. The tool isn't going to be winning points of phrenology here. Weight, hair color, and hairstyle would also likely tell you a lot.
I don't have any particular reason to believe this tool wouldn't work, but let's not pretend it's getting their by phrenology-esque topologies of people's faces.
A randomly chosen black individual in the united states has a > 72% chance of leaning democrat. A randomly chosen hispanic individual is ~55-65% chance of leaning democrat. I don't find it crazy to imagine they've got a few other smaller features to boost it.
> The accuracy is expressed as AUC, or a fraction of correct guesses when distinguishing between all possible pairs of faces—one conservative and one liberal.
I've never seen something like this. Maybe this is a normal procedure?
But I would be worried that the number of old black conservative women would be really small. Seems a bit sketchy
Simplistically, let’s take the above statistic “A randomly chosen black individual in the united states has a 72% chance of leaning democrat” at face value. So, a coin flip would be lower than 50-50 because someone of that race in that country does not have a 50 50 chance. So you would adjust the chance to 72-28 and compare that to the Facial recognition results. If you find that the results are the same, then you know that the Facial Recognition not picking up on anything beyond race. If the results are different, you know the FR is picking up on something in addition to race.
Really it is more complex than that, but fundamentally you try to say “how accurate can we be using just age, gender, and ethnicity” and use that as your controlled benchmark.
I understand what they're implying by "adjusted accuracy". My point is that I'm not sure that metric really makes sense, because "accuracy" isn't a particularly useful metric to begin with. It depends entirely on the sample distribution. "Always guess not fraud" will be 99.9% "accurate" for most use cases.
I'm asking what the literal metric is.
edit: and I don't think your explanation really works for accuracy, because accuracy isn't a relative measure, like, say, R2.
Did it? If you control for gender but not sex, you can use the difference to predict ideology. And for ethnicity, there are subethnicities that matter too - white Italian and white German have different proclivities.
72% is a very significant deviation from 50% though, I wasn't expecting such a result.
>The highest predictive power was afforded by head orientation (58%), followed by emotional expression (57%). Liberals tended to face the camera more directly, were more likely to express surprise, and less likely to express disgust.
Emotional expression makes some sense in hindsight but I wouldn't have though that head orientation would correlate. It's interesting to know how we betray ourselves with these minute details of body language.
Those categorizations are opposites on a continuum of various kinds of muscle tension, notably brow scrunch muscles. Surprise is literally brow up and jaw slack, disgust involves brow scrunch and lip tightening. Facial expressions also bleed through to our experienced emotions, so going around with face scrunched a lot will MAKE you more suspicious and disgusted with things.
A 72% score on this scale is equivalent to confidently knowing 44% of the answers, and coin-flipping the rest.
It's doing something that untrained humans are not capable of [edited to add: although humans were apparently tested by a different method, so this is not properly comparable], but is still a failing grade by usual methods of assessing human knowledge.
So... now, is it because people that came from the same regions a long time ago all happened to tend towards a certain political leaning and that just has stayed within families for that long?! Like a culture thing, but within families?
In general, though, its amazing how little control we have over who we are despite the feeling that we are in control of it.
Could be but not across the board true. My parents are deeply liberal to the point of the Constitution be damned so long as their views are enacted in some way. I am just the opposite - a constitution conservative. I am aware of several friends families that are the same or opposite.
I think the key to you last comment is how much intellectual control someone has over their emotional response. I like to think I am pretty good with that at this present time. That was not always the case and may not always be the case.
So, unless I read the abstract wrong, this is identifying age, race, and gender, then making categorical qualifiers based on those distinctions. While using ML to do the facial recognition to distinguish a person's age, race and gender is neat - categorizing their political affiliation with a 72% accuracy rate is fairly nominal, given the tools used by modern parties to garner donations and directed online advertising - no offense.
I don't really buy this study. If humans can't get a better accuracy than 55%, I'm convinced this vector is leaking some obvious (maybe high-frequency) information that doesn't have anything to do with the face. E.g location of the person.
> Allport and Kramer (1946) randomly presented 20 yearbook photographs of Jews and Non-Jews to 223 undergraduate students for 15 s each and asked them to categorize the
person in each photograph as Jewish or non-Jewish, or to
pass on the trial by indicating a lack of knowledge. The
reported median identification for the sample was slightly
above chance (55.5%; Allport & Kramer, 1946). Moreover,
they found that highly prejudiced people were more accurate
at distinguishing Jews from non-Jews
So it's not the same set of photos and not the same question.
When I read a paper like this I'm looking for four things: (1) the data, (2) the benchmarks, (3) the architecture, (4) the controls/ablation.
1. The data:
"We used a sample of 1,085,795 participants from three countries (the U.S., the UK, and Canada; see Table 1) and their self-reported political orientation, age, and gender. Their facial images (one per person) were obtained from their profiles on Facebook or a popular dating website... Facial images were processed using Face++37 to detect faces. Images were cropped around the face-box provided by Face++ (red frame on Fig. 1) and resized to 224 × 224 pixels."
2. The benchmarks:
"For example, when asked to distinguish between two faces—one conservative and one liberal—people are correct about 55% of the time."
3. The controls:
"What would an algorithm’s accuracy be when distinguishing between faces of people of the same age, gender, and ethnicity? To answer this question, classification accuracies were recomputed using only face pairs of the same age, gender, and ethnicity."
A. A complaint:
Geography and income are two powerful conditioners. These can leak in so many ways: uncropped background (geography), image color and quality (income), eyeglass shape (geography and income). This study really needs more controls. Geography and income would be a nice start.
Geography and income are two powerful conditioners. These can leak in so many ways: uncropped background (geography), image color and quality (income), eyeglass shape (geography and income). This study really needs more controls. Geography and income would be a nice start.
But then the data wouldn't represent the natural world: nature as it is.
Raw data is the correct thing to use, because it's what a hypothetical other person would also use if you ran the same experiment yourself.
That wasn't the claim. The claim here is that we should scrub certain faces from the dataset in order to change the dataset in a certain favorable way.
Yep, these papers don't usually pass the sniff test. My bet is you can predict the phone brand from the camera grain and that correlates with geography & income.
This reminds me of an early ML study about detecting skin cancer from pictures with a high accuracy rate.
The problem was, that with the ML, they ended up building a ruler classifier, because most of the pictures with skin cancer happened to also have a ruler in them to measure the size.
Or the commercial model that identifies criminals from their photograph. Turns out people who frown are criminals. People who smile aren't. Or so you'd believe if you anchored your expectations comparing mug shots to social media profile pictures.
My political leanings change based on how recently a drank a cup of tea. This obsession with a left-right spectrum is a big part of the problem now.
You think teachers are underpaid? Oh obviously you must be a pro-abortion, $15 minimum wage supporting, transgender-rights activist.
What's that you say, Christian bakers should be allowed to refuse to bake a cake with a pro-gay message on it? Oh, you must be a gun-toting, pro-life, anti-immigratnt Trump fanatic.
This kind of sorting people into simple binary categories, and giving them a "shopping bag" full of opinions they're supposed to hold helps nobody.
I'm not sure how this was relevant in anyway to your comment, but I just kinda jumped on my soapbox there.
Apart from the 55% human accuracy, which apparently comes from a completely unrelated study, the bit that really stands out to me is it reports the accuracy only drops from 72% to 68% when controlling for demographic accuracy in the US (a little more noticeable in the UK). Considering demographics alone gets you 60-90% accuracy on voting intention for many US demographics, it strikes me as extremely odd it would have have so little
impact on the model.
I believe that these descriptors are created only based off the visual image:
A face descriptor is obtained from the learned networks as follows: the centre 224 × 224 crop of the face image is used. The shorter side is resized to 256, and the CNNs descriptor is computed for this region by extracting the deep features from the layer adjacent to the classifier layer. This leads to a 2048 dimensional descriptor, which is then L2 normalised.
> Both in real life and in our sample, the classification of political orientation is to some extent enabled by demographic traits clearly displayed on participants’ faces. For example ... white people, older people, and males are more likely to be conservatives. What would an algorithm’s accuracy be when distinguishing between faces of people of the same age, gender, and ethnicity? To answer this question, classification accuracies were recomputed using only face pairs of the same age, gender, and ethnicity ... The accuracy dropped by only 3.5% on average
From there, it would seem that cues might come from how 'kempt' they appear, whether the head shot was from a party or for a resume, perhaps color and style of clothing, ... I.e., maybe not strictly the face.
If the total population sampling is the same, I would expect the accuracies to remain the same. E.g. if I can get 72% accuracy in the total population just by looking at age/race/gender, doesn’t that exactly mean the accuracies in each individual category are on average 72%?
Not necessarily because each age/race/gender tuple can be present in the test dataset different amounts, and either be a stronger or weaker indicator to the model.
I suspect there are a lot of less obvious things they'd need to control for. Off the top of my head, weight would be an obvious one; in developed countries urban areas (particularly large urban areas) generally have a lower average BMI than rural and suburban areas, and there's also typically a major political difference between rural and urban areas.
But wouldn't this be a reasonable feature used by the classifier to reach its conclusion? They can't control for everything, it would become meaningless.
I think the questions about age/sex/ethnicity are sensible in that it's a valid question to ask whether it's just doing the naive/obvious thing or something more. But if you keep on removing the less obvious things then of course you'll reach a point where it's no better than a coin flip because it's basically comparing blank pictures.
If only the researchers were as smart as you apparently think you are and somehow remembered to control for these very obvious factors. Oh yeah, they did.
I was not commenting on the main article about the research, but instead on the comment by user karaterobot that I am responding to which -- if I am interpreting it correctly -- asks the question how good one can guess purely based on those 3 demographic axes.
As you note, the researcher's predictor appears to do better than random even controlling for these obvious demographic skews, which is fascinating.
Actually, you have to answer something more from the other side: for a randomly picked person (voter), i.e. considering the distribution of race, gender and age in the population: how likely is a guess just based on these factors correct. The number you come up with might actually not be as far from 50% as you would expect.
This is the exact question I came to the comments to find.
The abstract states:
>Accuracy remained high (69%) even when controlling for age, gender, and ethnicity.
To give some context, chance is 50%, human guess is 55% and a 100-question questionnaire is 66%.
Personally, I am surprised that the accuracy remained that high when controlling for the three variables I would have considered most telling in the determination (age, gender and race).
I'd be very curious to know what exactly the algorithm is determining from the face photos outside of those obvious variables. I know with a ML algorithm it's practically impossible to determine why the classification was made, but does anyone human here have any thoughts?
"The dating website sample was provided by a popular dating website in 2017. It contains profile images uploaded by 977,777 users; their location (country); and self-reported political orientation, gender, and age."
It doesn't sound like people were consenting/aware that they'll endup on a facial recognition study.
But I'm not even surprised dating website would sell their such data
Ha. Hahaha. I wish. I'm sorry to laugh, but a ton of ML papers are based on illegally-scraped datasets of one form or another, unless they use strictly blessed datasets (Imagenet2012 being the gold standard mostly-useless-in-the-real-world dataset).
OpenAI's Jukebox is based on illegal large-scale gathering of copyrighted material, for example.
I've never encountered this term. I can see how scrapping might be a violation of some websites terms of use, but I've never seen "scrapped illegally" used. Do you have any examples?
Well, pictures of faces could be considered personal data per GDPR. Scraping that data without each person's approval could be illegal regardless of any terms of use.
- I have agreed with LinkedIn that they may use my personal information for a set of well-defined uses (basically things on the LinkedIn website/service, and some 3rd party services they use to run the website/service).
- LinkedIn promise that they will not share my identifiable personal information with 3rd parties for any use
- LinkedIn's terms of use state that nobody may scrape personal information from their website without their consent. This is how they enforce the previous promise to me
- Some business comes along and scrapes my personal information for their own business use.
- That business knows that LinkedIn prohibit this, and they know that I have only consented for my personal information to be used for LinkedIn itself.
- This is probably "unlawful" (as they're interfering in my contract with LinkedIn), and certainly violating my GDPR rights. Sadly, it's hard to point at a specific example as guidance doesn't have a section titled "Can I ignore individual's explicit opting out of my usage?".
Hence, illegal scraping, as willing violating the GDPR is illegal.
Just to head-off the very common response: Personal, individual, use is not covered by the GDPR. So there is nothing wrong with you going and using my LinkedIn data for any personal reasons. The moment you try to use it for business purposes though, that's illegal.
Exactly. There was some fuss a while back about a similar classifier for sexuality. It turned out to be guessing mostly based on head tilt, personal hygiene and whether the person was wearing glasses. The physiognomy component was ~nonexistent even though it was publicized as though it weren’t. People intentionally if at times subconsciously present themselves in a way that signals information to kindred spirits. You’d need to bring in hundreds of people, wash them and basically take mugshots to control for that.
People don't get confirmation of those details normally unlike the ML algorithm. Of course we don't lock people in boxes with a stack of training set photos for a period of time equivalent to ML training.
At least partly due to lack of feedback on accuracy. I don’t know about you but I don’t necessarily ask everyone I meet their political leanings, so it’s hard to train yourself other than through stereotype.
My guess would be sampling bias. Most people base their model on small geographically restricted samples, and are heavily biased by media. It is possible that some people can repeatedly perform better than 55%.
One can only wonder what the result would be by using the equivalent of government ID photos (neutral expression, no smile or make-up, solid backgrounds).
The problem wit this is bad interpretation, people look at it and think oh this means political orientation is somehow inherent to biology or some such nonsense. Any data that can make predictions is just a sign that in this data there are some patterns, but there is never a straightforward interpretation for it.
Also, whoever makes the first app for this will go viral 100%.
Surprising, yet not surprising. As the article mentions:
> Both in real life and in our sample, the classification of political orientation is to some extent enabled by demographic traits clearly displayed on participants’ faces. For example, as evidenced in literature and Table 1, in the U.S., white people, older people, and males are more likely to be conservatives.
Most people can predict a person's political orientation of their own country with >50% accuracy as well just by looking at a face. Black or latino? probably liberal. Old white person? probably conservative. If you can see more than just their face it's even easier (wearing religious paraphernalia? LGBT paraphernalia? etc?)
What I thought was interesting was:
> The algorithm could successfully predict political orientation across countries
I was under the impression that "liberal" and "conservative" had different meanings in UK vs. USA so how could it do this?
I am curious how much additional information, if any, is learned from a person's face. From a full-body photo, with the face blacked out, what can be learned?
I haven't read the paper yet, but it's not that surprising. I think it's about sociology more than biology.
In the US, African Americans vote overwhelmingly for democrats for example. Skin color therefore becomes a very good predictor of political orientation. You can probably extend this to states being populated from various migration waves, say 'people who look like Danes vote for republicans because state X was populated by Danes and votes Republican' . Carry this across generations/education, and you may have an explanation.
> Their facial images were obtained from their profiles on Facebook or a popular dating website
So, this seems to be doing profile picture recognition, not facial recognition. It's not like they're saying there are physical features in your face that give away your political affiliation - that is to say, putting two people's bodies under the same photographic conditions would probably not create this kind of signal.
What this is probably training on is the cues for cultural values that we self-select in our most deliberately promoted images of ourselves. If you have a carefully-chosen profile picture, it probably includes signals of what's important to you, especially if you're using it to attract people with similar values.
still leaves the micro- and macro-expressions, small grooming cues (makeup, no makeup, eyebrows trimmed or not), hairline, head angle vs. camera, lighting etc. These are all things that humans very specifically deploy to define themselves and their grouping, and communicate with others. So I am guessing a whole universe of personal yes-no qualities, political and otherwise, are encoded there, quite intentionally.
I'm sure you're right. But it makes no difference, because when the facial recognition is applied, I think the likelyhood that people will have the same "grooming cues" as they do in their profile pictures is pretty high. If I have short hair in my profile pic, I probably have short hair in my everyday life, in the moment that my face is recorded and processed, as well. Yes, I choose how my face looks myself. But that does not mean anything, as objectively it is the same, whether I intended to or not.
I assume there's no way to actually verify how someone may choose to vote, assuming there's no record of that?
I think there's huge value now that everything is being sent into a "machine" or "the algorithm" in fucking with it.
Order sex toys from Amazon, show them you're into outrageous books and fool them into creating a fake profile of "you", based on your spending, browsing and other data you generate.
I'd love to ask a machine what it knows about me, how accurate it is, and then switch it all up. I'm too old to vote now (and will probably be dead soon) but I'd love to pick a position completely unexpected just to throw it off.
From Amazon's perspective you have not poisoned the well... you are someone who is likely to buy those things. Amazon does not care if you really like or even use them, it only cares if you buy them.
> "Accuracy remained high (69%) even when controlling for age, gender, and ethnicity."
So if I just assign the majority label to all of the population of a given demographics group, I would get the same result right? i.e., predicting "left" for all minorities under 30. You would also get ~70% accuracy.
You quoted the part about controlling for age, then described the sort of mistake that comes from not controlling for age. So I would like to know what you think it means, in order to meet you where you are. I'm not expressing aggression toward you.
My interpretation of control is fixing all other variables (the ones they mentioned) except for the one being measured (political orientation). If that's not what they did I'm happy to learn.
In that case I don't understand your original comment, as it describes the sort of mistake that arises when you don't control for other factors, but you appear to accept that they did.
> So if I just assign the majority label to all of the population of a given demographics group, I would get the same result right? i.e., predicting "left" for all minorities under 30. You would also get ~70% accuracy.
I meant that even if you control age, gender, ethnicity, a very trivial predictor (i.e., always predicting the majority label) could yield similar performance. What I meant to say was that their model may not perform as well as they made it sound.
420 comments
[ 3.9 ms ] story [ 315 ms ] threadhttps://www.sciencedirect.com/science/article/pii/S235215462...
https://onlinelibrary.wiley.com/doi/abs/10.1111/pops.12230
I'm sure many more subtle examples also exist.
1: https://de.wikipedia.org/wiki/Habsburger_Unterlippe
> Accuracy remained high (69%) even when controlling for age, gender, and ethnicity.
Some people might be better at some things, while others are better at other things. (thus combining strengths)
Some people might actually be really bad at some other things yet again, while others could happen to be very good at them (thus covering each others weaknesses)
When you have people with different backgrounds and training all working together you can do things together that you wouldn't otherwise be able to do separately.
But still, it's yet another facet of our lives where we deal with disagreements by putting them out of sight rather than communicating and understanding.
I tend to agree. It's why I find all of the recent "Diversity is our strength" stuff so puzzling.
I'm not against diversity, I just recognize that it generally results in at least some interpersonal challenges to overcome, not necessarily unity.
https://en.wikipedia.org/wiki/E_pluribus_unum
https://en.wikipedia.org/wiki/Melting_pot
In practice, the multicultural empire was greatly weakened by incessant nationalist bickering.
Once you have to declare that X is your strength or something similar, it most likely isn't.
I would argue that Diversity is necessary, but not sufficient.
("E pluribus unum" is -of course- the motto of the United States of America. Not the least of nations!)
"If we all reacted the same way, we'd be predictable, and there's always more than one way to view a situation. What's true for the group is also true for the individual. It's simple: Overspecialize, and you breed in weakness. It's slow death. "
Multiculturalism is dead.
In all seriousness, I haven’t followed closely, so I don’t know who hates Dr. Seuss lately, they are the baddies.
That initial idea somehow turns into, "The left hates Dr. Seuss" and wild slippery slopes about book burning.
https://sophia.stkate.edu/cgi/viewcontent.cgi?article=1050&c...
Recently the publisher announced it was going to stop selling those - I think it was 6 books out of 107 in the catalog. There was no pressure campaign or wave of outrage driving it.
Perhaps you should choose better news sources, as the ones you are using seem to be serving you poorly.
The outrage over racial overtones came from a left aligned organization. Fact.
It was carried on and cheered on by left leaning people. Fact.
People on the right don’t use words like “problematic” and “racialized undertones”
That’s on you guys.
This is from The horses mouth:
“ Six popular Dr. Seuss books — including And to Think That I Saw It on Mulberry Street and If I Ran the Zoo — “will stop being published because of racist and insensitive imagery, the business that preserves and protects the author’s legacy said Tuesday,” The Associated Press reported.
“These books portray people in ways that are hurtful and wrong,” Dr. Seuss Enterprises told The Associated Press in a statement marking the late author and illustrator’s birthday.”
^ The above is just stupid.
However, I just shared a bunch of open source facts with you. All sourced from left leaning sources as to be somewhat unbiased (even though my bias on this subject is probably pretty heavy)
This isn’t about winning an online argument, it’s about a call for reason when things have become so silly that we are a step away from the Grinch being canceled because his big nose looks Jewish and that’s Nazism or something.
Not one single neo Nazi has ever been uplifted and supported in their beliefs by classic children’s books.
On the right you have Trump cultists who loyally cling to every word he utters, regardless of it is true or not, spreading lies and misinformation on social media about voter fraud, vaccines, etc. You also have covid denial on the right where "the constitution" is used as a bludgeon to resist any reasonable public health policy like mask usage.
On the left you have the wokeness mobs that go around trying to cancel/censor/destroy every historical figure/book/statue who didn't live a perfect life and destroying careers of anyone who says the wrong word or makes a bad analogy. You also have covid zealotry on the left where "covid deaths prevented" is prioritized above all else and we can't re-open schools, re-open businesses, visit grandparents, or otherwise get back to normal life (even after mass vaccination) until 100% of all remaining unknowns are known (even if it takes another 2 years of masking, zooming, and hermiting).
These are all very recent examples.
So it would be interesting to see if hyperpartisans could be separated from moderates by their appearance (just to be clear, I dont believe in phrenology, I imagine its information leaking from other aspects of presentation and background).
With respect to humour, the modern lack of humor seems to focus on not offending people, which I admit I dont understand but I could guess is predicated on the feeling that laughter is somehow equated to divisive mocking, as opposed to fun. I bring this up only because it reminds me a lot of Umberto Eco's "The Name of The Rose" where the 13th century religious leaders were arguing that laughter was inappropriate, and Jesus never laughed, somehow rooted in the belief that finding humor in things admitted the possibility of laughing at aspects of religion and therefore not taking them seriously.
Anyway, your comment made me think, so thanks.
This sounds more like people who are from minority groups (or who look like they are, to a computer vision algorithm) are more likely to agree with left-leaning policies, which probably has more to do with the policies in those countries than it does with any sort of genetic features. I feel like this might not work as well in non-Western countries, for example.
I did not do the full math but just the ballpark numbers I entered in to my calculator says that it should account for about 11%.
* 89% according to numbers from nbc
Edit: So say I assume that every African American person I see votes left. 13% of the population is African American 89% of them actually vote left. 13% * 89% = 11% Then I simply guess on all non African American a 50/50 shot. I should then be right 50% + 11% = 61% of the time.
Not sure I’m entirely comfortable with that.
The issue is potentially being included in a study unintentionally, that is impossible to anonymize, without prior knowledge or consent.
> Political orientation was correctly classified in 72% of liberal–conservative face pairs, remarkably better than chance (50%), human accuracy (55%), or one afforded by a 100-item personality questionnaire (66%).
The real question is whether the tool can beat a lookup table of age, race, and gender probabilities. The tool isn't going to be winning points of phrenology here. Weight, hair color, and hairstyle would also likely tell you a lot.
I don't have any particular reason to believe this tool wouldn't work, but let's not pretend it's getting their by phrenology-esque topologies of people's faces.
A randomly chosen black individual in the united states has a > 72% chance of leaning democrat. A randomly chosen hispanic individual is ~55-65% chance of leaning democrat. I don't find it crazy to imagine they've got a few other smaller features to boost it.
> Accuracy remained high (69%) even when controlling for age, gender, and ethnicity.
> The accuracy is expressed as AUC, or a fraction of correct guesses when distinguishing between all possible pairs of faces—one conservative and one liberal.
I've never seen something like this. Maybe this is a normal procedure?
But I would be worried that the number of old black conservative women would be really small. Seems a bit sketchy
Really it is more complex than that, but fundamentally you try to say “how accurate can we be using just age, gender, and ethnicity” and use that as your controlled benchmark.
I'm asking what the literal metric is.
edit: and I don't think your explanation really works for accuracy, because accuracy isn't a relative measure, like, say, R2.
>The highest predictive power was afforded by head orientation (58%), followed by emotional expression (57%). Liberals tended to face the camera more directly, were more likely to express surprise, and less likely to express disgust.
Emotional expression makes some sense in hindsight but I wouldn't have though that head orientation would correlate. It's interesting to know how we betray ourselves with these minute details of body language.
It's doing something that untrained humans are not capable of [edited to add: although humans were apparently tested by a different method, so this is not properly comparable], but is still a failing grade by usual methods of assessing human knowledge.
In general, though, its amazing how little control we have over who we are despite the feeling that we are in control of it.
I think the key to you last comment is how much intellectual control someone has over their emotional response. I like to think I am pretty good with that at this present time. That was not always the case and may not always be the case.
So they had more than age, race, and gender, but it doesn't really say how things were weighted.
Doesn't have to be that way. It could be age, age+1, age+2 ....
The only relevant 55% that I can find is:
> Allport and Kramer (1946) randomly presented 20 yearbook photographs of Jews and Non-Jews to 223 undergraduate students for 15 s each and asked them to categorize the person in each photograph as Jewish or non-Jewish, or to pass on the trial by indicating a lack of knowledge. The reported median identification for the sample was slightly above chance (55.5%; Allport & Kramer, 1946). Moreover, they found that highly prejudiced people were more accurate at distinguishing Jews from non-Jews
So it's not the same set of photos and not the same question.
1. The data:
"We used a sample of 1,085,795 participants from three countries (the U.S., the UK, and Canada; see Table 1) and their self-reported political orientation, age, and gender. Their facial images (one per person) were obtained from their profiles on Facebook or a popular dating website... Facial images were processed using Face++37 to detect faces. Images were cropped around the face-box provided by Face++ (red frame on Fig. 1) and resized to 224 × 224 pixels."
2. The benchmarks:
"For example, when asked to distinguish between two faces—one conservative and one liberal—people are correct about 55% of the time."
3. The controls:
"What would an algorithm’s accuracy be when distinguishing between faces of people of the same age, gender, and ethnicity? To answer this question, classification accuracies were recomputed using only face pairs of the same age, gender, and ethnicity."
A. A complaint:
Geography and income are two powerful conditioners. These can leak in so many ways: uncropped background (geography), image color and quality (income), eyeglass shape (geography and income). This study really needs more controls. Geography and income would be a nice start.
But then the data wouldn't represent the natural world: nature as it is.
Raw data is the correct thing to use, because it's what a hypothetical other person would also use if you ran the same experiment yourself.
That way we can see how much of the performance is from magic AI pixie dust, and how much is from basic 19th century statistics.
Every time I read a paper like this, I have this Margaret Mitchell talk [1] in the back of my mind.
[1] https://youtu.be/XR8YSRcuVLE
The problem was, that with the ML, they ended up building a ruler classifier, because most of the pictures with skin cancer happened to also have a ruler in them to measure the size.
> Their facial images (one per person) were obtained from their profiles on Facebook or a popular dating website
so of course the first thing to comes to mind is "how good of a predictor is just knowing which of those two sites the image came from?"
You think teachers are underpaid? Oh obviously you must be a pro-abortion, $15 minimum wage supporting, transgender-rights activist.
What's that you say, Christian bakers should be allowed to refuse to bake a cake with a pro-gay message on it? Oh, you must be a gun-toting, pro-life, anti-immigratnt Trump fanatic.
This kind of sorting people into simple binary categories, and giving them a "shopping bag" full of opinions they're supposed to hold helps nobody.
I'm not sure how this was relevant in anyway to your comment, but I just kinda jumped on my soapbox there.
A face descriptor is obtained from the learned networks as follows: the centre 224 × 224 crop of the face image is used. The shorter side is resized to 256, and the CNNs descriptor is computed for this region by extracting the deep features from the layer adjacent to the classifier layer. This leads to a 2048 dimensional descriptor, which is then L2 normalised.
https://www.arxiv-vanity.com/papers/1710.08092/
> Both in real life and in our sample, the classification of political orientation is to some extent enabled by demographic traits clearly displayed on participants’ faces. For example ... white people, older people, and males are more likely to be conservatives. What would an algorithm’s accuracy be when distinguishing between faces of people of the same age, gender, and ethnicity? To answer this question, classification accuracies were recomputed using only face pairs of the same age, gender, and ethnicity ... The accuracy dropped by only 3.5% on average
Though cropping can only do so much.
I think the questions about age/sex/ethnicity are sensible in that it's a valid question to ask whether it's just doing the naive/obvious thing or something more. But if you keep on removing the less obvious things then of course you'll reach a point where it's no better than a coin flip because it's basically comparing blank pictures.
People under 30: 60-36.
White men: 38-61
Black women: 90-9
So there are definitely some strong predictors there.
Source: https://www.businessinsider.com/2016-2020-electoral-maps-exi...
As you note, the researcher's predictor appears to do better than random even controlling for these obvious demographic skews, which is fascinating.
The abstract states:
>Accuracy remained high (69%) even when controlling for age, gender, and ethnicity.
To give some context, chance is 50%, human guess is 55% and a 100-question questionnaire is 66%.
Personally, I am surprised that the accuracy remained that high when controlling for the three variables I would have considered most telling in the determination (age, gender and race).
I'd be very curious to know what exactly the algorithm is determining from the face photos outside of those obvious variables. I know with a ML algorithm it's practically impossible to determine why the classification was made, but does anyone human here have any thoughts?
Could it be a version of this: https://hackernoon.com/dogs-wolves-data-science-and-why-mach...
It doesn't sound like people were consenting/aware that they'll endup on a facial recognition study.
But I'm not even surprised dating website would sell their such data
This dataset directly from the dating website
OpenAI's Jukebox is based on illegal large-scale gathering of copyrighted material, for example.
I've never encountered this term. I can see how scrapping might be a violation of some websites terms of use, but I've never seen "scrapped illegally" used. Do you have any examples?
- I have agreed with LinkedIn that they may use my personal information for a set of well-defined uses (basically things on the LinkedIn website/service, and some 3rd party services they use to run the website/service).
- LinkedIn promise that they will not share my identifiable personal information with 3rd parties for any use
- LinkedIn's terms of use state that nobody may scrape personal information from their website without their consent. This is how they enforce the previous promise to me
- Some business comes along and scrapes my personal information for their own business use.
- That business knows that LinkedIn prohibit this, and they know that I have only consented for my personal information to be used for LinkedIn itself.
- This is probably "unlawful" (as they're interfering in my contract with LinkedIn), and certainly violating my GDPR rights. Sadly, it's hard to point at a specific example as guidance doesn't have a section titled "Can I ignore individual's explicit opting out of my usage?".
Hence, illegal scraping, as willing violating the GDPR is illegal.
Just to head-off the very common response: Personal, individual, use is not covered by the GDPR. So there is nothing wrong with you going and using my LinkedIn data for any personal reasons. The moment you try to use it for business purposes though, that's illegal.
2) I haven't the slightest clue why you think scraped data can't be used in a publication - some results from first page of Google Scholar:
https://www.sciencedirect.com/science/article/pii/S187802961...
https://journals.sagepub.com/doi/abs/10.1177/004209802091819...
It's a study which shows that pictures that people select to represent themselves publicly have features that indicate political leaning.
Also, whoever makes the first app for this will go viral 100%.
> Both in real life and in our sample, the classification of political orientation is to some extent enabled by demographic traits clearly displayed on participants’ faces. For example, as evidenced in literature and Table 1, in the U.S., white people, older people, and males are more likely to be conservatives.
Most people can predict a person's political orientation of their own country with >50% accuracy as well just by looking at a face. Black or latino? probably liberal. Old white person? probably conservative. If you can see more than just their face it's even easier (wearing religious paraphernalia? LGBT paraphernalia? etc?)
What I thought was interesting was:
> The algorithm could successfully predict political orientation across countries
I was under the impression that "liberal" and "conservative" had different meanings in UK vs. USA so how could it do this?
I assume they're using the US definition (meaning "left wing" and "right wing", more or less).
In the US, African Americans vote overwhelmingly for democrats for example. Skin color therefore becomes a very good predictor of political orientation. You can probably extend this to states being populated from various migration waves, say 'people who look like Danes vote for republicans because state X was populated by Danes and votes Republican' . Carry this across generations/education, and you may have an explanation.
So, this seems to be doing profile picture recognition, not facial recognition. It's not like they're saying there are physical features in your face that give away your political affiliation - that is to say, putting two people's bodies under the same photographic conditions would probably not create this kind of signal.
What this is probably training on is the cues for cultural values that we self-select in our most deliberately promoted images of ourselves. If you have a carefully-chosen profile picture, it probably includes signals of what's important to you, especially if you're using it to attract people with similar values.
[0] https://imgur.com/FK0RzKM
https://i.kym-cdn.com/photos/images/newsfeed/001/330/806/8fc...
I think there's huge value now that everything is being sent into a "machine" or "the algorithm" in fucking with it.
Order sex toys from Amazon, show them you're into outrageous books and fool them into creating a fake profile of "you", based on your spending, browsing and other data you generate.
I'd love to ask a machine what it knows about me, how accurate it is, and then switch it all up. I'm too old to vote now (and will probably be dead soon) but I'd love to pick a position completely unexpected just to throw it off.
Poison the well
There are places with a maximum voting age?
So if I just assign the majority label to all of the population of a given demographics group, I would get the same result right? i.e., predicting "left" for all minorities under 30. You would also get ~70% accuracy.
> So if I just assign the majority label to all of the population of a given demographics group, I would get the same result right? i.e., predicting "left" for all minorities under 30. You would also get ~70% accuracy.
I meant that even if you control age, gender, ethnicity, a very trivial predictor (i.e., always predicting the majority label) could yield similar performance. What I meant to say was that their model may not perform as well as they made it sound.