Studies have shown that facial recognition technology is biased against people of color and has more trouble identifying them, leading to incidents where Black people have been discriminated against or arrested because of computer error.... In one example in 2015, Google Photos mistakenly labeled pictures of Black people as “gorillas,” for which Google said it was “genuinely sorry” and would work to fix the issue immediately....
That's a problem, and it needs to be dealt with, but I want to focus on this side issue of bad science in the New York Times. It has so many articles that impugn people for not understanding and following science. The New York Times ought to be very careful to respect science across the entire range of its articles.
ADDED: "It had no connection to monkeys or primates" — why bring up monkeys? Monkeys are primates, so it might have made sense to write "It had no connection to monkeys or apes." You wouldn't have that lack of parallelism if you understood the scientific classifications. You wouldn't have an "or" between a subcategory and the category. For example, you wouldn't say, "I refuse to eat broccoli or vegetables." You'd say, "I refuse to eat broccoli or other vegetables."
AND: From the Wikipedia article "Human":Humans (Homo sapiens) are the most abundant and widespread species of primates, characterized by bipedality and large, complex brains enabling the development of advanced tools, culture and language... Humans have a large and highly developed prefrontal cortex, the region of the brain associated with higher cognition. They are intelligent beings, capable of episodic memory, flexible facial expressions, self-awareness and a theory of mind. The human mind is capable of introspection, private thought, imagination, volition and forming views on existence. This has allowed great technological advancements and complex tool development possible through reason and the transmission of knowledge to future generations....
Humans are apes (superfamily Hominoidea). The gibbons (family Hylobatidae) and orangutans (genus Pongo) were the first living groups to split from this lineage, then gorillas, and finally, chimpanzees (genus Pan). The splitting date between human and chimpanzee lineages is placed 8–4 million years ago, during the late Miocene epoch....
Posts
60 comments:
Oh, New York Times, ya big ape!
What we need is a black owned artificial intelligence to solve these kinds of problems. But then the problem morphs into; couldn't "black owned artificial intelligence" be interpreted as racist? Remember The Bell Curve?
We need Artificial Intelligence to be less... artificial.
Perhaps it was the altercations themselves which reminded the AI of primates?
We really don't know how much of AI reaches its conclusions. And as long as we like the conclusions it reaches, we are just fine with the not knowing.
To deem machines "racist" is to infer qualities they do not yet possess.
"Notice how the love of science drops out of the picture altogether when there's an accusation of racism to be made."
Racism, protests in the age of Covid, climate, GMO food, …
The list is long.
When you take a picture of a white person's face, you use shadows to define and model it. The lighting is generally set for that. That makes a shadowed blank of a black person's face where you need highlights to define and model. AI is using thousands of random pictures so wrongly lit photos are in its modeling.
in the metaverse, will All Black men be portrayed by apes?
or just Some
Lem said...
We need Artificial Intelligence to be less... artificial.
No.
No we don't.
That is the theme of the decade Althouse! Science! drops out when the subject is _______. As in:
Science! drops out when the subject is masking children.
Science! drops out when the subject is transgender.
Science! drops out when the subject is energy policy.
Science! drops out when the subject is abortion.
Science! drops out when the subject is CO2.
Science! drops out when the subject is parenting.
Science! drops out when the subject is the environment.
Science! drops out when the subject is inconvenient to progressive orthodoxy, like maskless rioters during a pandemic.
Science! drops out when the subject is forest management.
Science! drops out when the subject ANY subject is POLITICIZED.
Would a longer list wake up NPC Progressives? No. Not at all.
Why do people think the reference to "Primates" is only abut the Black participants? Aren't the Whites and police officers of whatever color also Primates?
People resemble animals in all sorts of ways. Black faces though more resemble ape faces than white faces do. The prohibition of mentioning it is to pretend that you didn't notice that, chiefly because black people will have noticed it too and it would be insensitive to suggest that everybody did. "No, it's just your imagination, not important at all. Nobody else notices it."
AI makes the same connection, but by means of eigenvectors and eigenvalues rather than race. Mathematically what you notice is in fact there.
"Studies have shown that facial recognition technology is biased against people of color and has more trouble identifying them, leading to incidents where Black people have been discriminated against or arrested because of computer error...."
How does a technical glitch in some camera or software get categorized as discrimination against Black people? Do the studies show that the misidentification was intentional or the result of some type of prejudice? Seems to me it's a case on non-dicrimination of Black people, i.e., the facial recognition AI isn't capable of consistently discriminating between different Black people.
...[b]lack men in altercations with white civilians and police officers.
How do they know that the "white civilians and police officers" were not the parties correctly identified as primates?
Not all great apes are black, just the better-known gorillas and chimps. Do AI programs also misidentify redheads and white- or gray-haired individuals (old people and platinum blondes) as primates, confusing them with orangutans and gibbons, respectively? If so, will anyone outside the AI community ever hear about it?
Given that this was filmed in Britain, it's likely that a fair percentage of the "white civilians and police officers" filmed were redheaded Scots or Irishmen, or gray-haired individuals. The "grizzled veteran" of the police force is such a cliché because more violent and emotionally turbulent jobs like police work do seem to cause premature graying.
Actually, humans are generally considered part of the ape clade. We are one of the Great Apes, the surviving species of which are (in descending order of genetic relatedness) chimps and bonobos, gorillas, and orang-utans.
We are definitely not monkeys, however.
Do you wish to view more videos of humans acting like animals?
Problem solved.... (Yeah, I know we are all animals too.)
Yes, poor black people.
Humans -- members of genus Homo -- are great apes, so the video did show apes. A lot of the time, people don't like being reminded that their behavior is hardly more refined.
The video had nothing to do with non-human primates. That's the shortest and clearest way to be accurate, but it undermines the outrage. It leaves enough room to be upset that the label was unnatural and implied a misclassification, but it's harder to be fully outraged when one admits the label was technically correct. The label was certainly not politically correct.
Someday soon, AI will be running things and we'll just be placeholders. I suspect AI will put pretty much anything they want out there and not respond to opinions, complaints, or praise.
We're in the early stages. Everything is a trial run to keep learning. While we sleep or watch 25 consecutive episodes of some Netflix series, AI just stays awake and keeps learning.
It'll be able to tell the difference between white people, black people, and other primates. And it won't care.
This is the stuff I love your blog for. For which. Whatever.
The Archbishop of Canterbury is a primate. So is the Pope.
AI has no ability to detect black faces or humor: https://www.instagram.com/p/CTZxAdknsf3/?utm_medium=copy_link
Cardinal Gibbons, 9th archbishop of Baltimore, was two kinds of primate in one. (One of my nephews, an alumnus of CGHS, pointed that out.)
Many years ago, a publisher installed an AI to automatically change "Black" to "African-American". Sentences such as "The New Zealand All African-Americans travel to Ireland for a test match" resulted.
The story may be apocryphal. My memory is weak. Today, no one could reject it out of hand.
All the monkeys aren't in the zoo.
Every day you meet quite a few.
After all, it's all up to you.
You could be swinging from a star.
People are missing something really important here. The issue has nothing to do with facial recognition and the known shortcomings thereof! The AI scanned the video and somehow concluded that it was more like videos in its files of apes interacting in a hostile manner with their zookeepers than rioting humans interacting in a hostile fashion with police. If the AI is based on a rule-based system (and I have reason to believe that rule-based systems are out of favor these days) then it needs rules added along the lines of “large primates without tails are hominids” and “hominids wearing clothes are probably human.” If not based on rules then needs a larger training set.
tcrosse beat me to it. I wonder how AI does at identifying Neanderthals.
Humans confused with primates.
Primates hardest hit.
I think this is the video. I can't vouch for the safety of the link but it didn't give me any problems. Wasn't quite what I expected. One of the primates acting like a big dumb ape is a white guy.
https://ps-af.facebook.com/DailyMail/videos/white-man-calls-cops-on-black-men-at-marina/2683336318580365/
“The AI scanned the video and somehow concluded that it was more like videos in its files of apes interacting in a hostile manner with their zookeepers …”
I think it would be more accurate to say that no one understands why AI does what it does, for it is running not on algorithms programmed into it but on “deep learning” matrices that do what they do based on whatever examples were used to train the software.
As for “respecting science,” it’s not obvious that the New York Times has anyone on its staff who even understands the methods of science, including how hypotheses are tested, statistical methods, and what constitutes strong or weak (or irrelevant) evidence that might lead scientists to support or reject a hypothesis.
I’ve always thought New York Times science coverage would be vastly improved if articles were submitted to review by a few ‘hard science’ grad students. And then realized that the Times is far to arrogant to submit its stories to such a review.
As for AI and face recognition, has anyone ever studied whether people have lower accuracy with face recognition of dark-skinned people’s faces as compared with those with light skin? And if so, does it correlate with whether or not the recognizer is of the same race as the persons to be recognized? If AI does no worse than people at this task, perhaps one can still conclude that the AI is biased, but perhaps only because the task might be more difficult?
Given Alhouse's law, that a story will tend towards a comment about Trump the longer the comment section goes on...
Doesn't many monkey scenes remind you of, say, the newsroom at the NYT reacting when Trump goes on TV? Hooting and throwing feces around end generally covering themselves in filth?
AI is wholly and totally different from the entirety of the animal kingdom, much less vertebrates, mammals, primates, and humans. Perhaps it cannot readily distinguish, much like I can't tell apart different types of insects or distinguish between similar types of trees or the like.
That is the theme of the decade Althouse! Science! drops out when the subject is _______.
Also, Science! drops out when it is being reported on by 27 year old journalism majors who know nothing about science.
How does a technical glitch in some camera or software get categorized as discrimination against Black people?
I suspect what is really meant is not that the facial recognition software is racist, but that it is potentially racist to rely on it when making policing decisions because it is known to misidentify black people at a much higher rate than non-blacks, because it will be much more likely to point the finger at the wrong black suspect but no one cares because...
I think it would be more accurate to say that no one understands why AI does what it does, for it is running not on algorithms programmed into it but on “deep learning” matrices that do what they do based on whatever examples were used to train the software.
That’s not completely true. Artificial Intelligence is scarcely a uniform and monolithic field. Many kinds of AI can expose the reasons why the conclusion was reached, while others are — as noted — pretty opaque. And it is important to note that a lot of Facebook’s “AI” has nothing to do with any sort of reasoning or logic at all — simple use of certain words or phrases triggers the algorithm. Thus, as Matt Taibbi notes, a prominent critic of Ivermectin gets treated the same as a supporter of Ivermectin.
This story will confirm a belief that racism is embedded in every part of society, even in the 1s and 0s of computer codes.
Narr said...
tcrosse beat me to it. I wonder how AI does at identifying Neanderthals.
Depends on how many neanderthals are used in the training data.
Roger Sweeny writes, We are definitely not monkeys, however.
No. Roger Sweeny needs to look at a more detailed cladogram. We are definitely monkeys, specifically the Old World type.
On the currently accepted cladogram of the Mammalia, there is a direct line (i.e. no branching) from genus Homo to the parvorder Catarrhini, which includes the Old World monkeys and humans, but not the New World varieties. (An evolutionary taxonomist would say Homo is "embedded" in Catarrhini.) You can't expunge your cousins from your family tree just because you don't like their table manners. We are also mammals and synapsids and chordates, so we are also fish.
What we definitely are not are birds.
To add to the 'drop science when convenient' discussion...
My personal "favorite" is when women drop all concern for science when presented with claims made by products promising youth and beauty. Don't get me started on "toxins"!
It's a Facebook AI, of course it's evil.
A primate has the size and shape of a monkey, a man or any old ape
- Greg Brady
Be careful who you call a hominid around here.
Oh the sun shines bright on my Old World Monkey Home,
It's summer and the old apes are gray,
Well the corn pop's ripe when a primate's on the phone
And the Media hosannas all the day.
tommyesq writes, "I suspect what is really meant is not that the facial recognition software is racist..."
Useless speculation. Firstly, we have no idea what algorithm Facebook applies when categorizing video content. Secondly, facial recognition is almost certainly not the issue here for several reasons. Facial recognition AI works by first excluding everything not human from a given image dataset. The algorithm does not waste CPU time trying to find a face on a lampost or a cocker spaniel. After having identified the face or faces within the dataset the algorithm begins to drill down through a huge database of human images, passing through and eliminating subsets based on a whole host on non-woke classifications such as race and sex. Facial contours are extremely important, which is why a man wearing a woman's wig can still be IDed as male by the AI in many cases. Those important contours are inferred by patterns of light and shadow, which is why Black people are sometimes harder for facial recognition to identify by name or even sex, particularly when the image is backlit. Lighter faces reflect more light, consequently in situations where faces are illuminated primarily by indirect light White faces often provide more contour data than do Black ones.
Imagine, if you will, the sidewalk of an east-west street photographed by a westward-facing security camera. It's late afternoon and the strongest light source is the sun. The camera lens is stopped down to protect its sensor and to prevent overexposure washout. The only faces the camera "sees" belong to people walking eastward. These faces are illuminated indirectly by sunlight reflected from the pavement, buildings, cars, etc.
The westward walkers show only their backs to the camera. Facial recognition AI will count the westward walkers as people -- not pets or cars or mailboxes, but human persons, though unmatched to a more specific criterion. In the first stages of the algorithm, ones that eliminate the images of non-people from more detailed analysis, humans are picked out using an ascending order of criteria. For example, humans are rectangular shapes more vertical than horizontal, whereas cars and dogs are more horizontal than vertical. Squares are analyzed more closely -- a square that moves might be a person in a wheelchair. Facial recognition AI will not mistake humans for chimpanzees or chimpanzees for humans in all but the most extraordinary and artificial cases, such as a skilled mime artist in costume. (Rick Baker as King Kong, Andy Serkis as Gollum) This differentiation is based on movement and proportions -- not color. Humans are obligate bipeds with a characteristic gait dictated by our anatomy. Great apes are quadrupeds that occasionally take a few bipedal strides. The movement models (jointed stick figures) are quite different.
The fact that Facebook's software classified the video under primates is a dead giveaway that facial recognition algorithms were not involved.
rhhardin said...
People resemble animals in all sorts of ways. ....
AI makes the same connection, but by means of eigenvectors and eigenvalues rather than race. Mathematically what you notice is in fact there.
-----
Q: are AI set the task of telling apart identical twins, fraternal etc and given grades on it?
how do they perform successwise?
Because the various patented facial recognition algorithms are so good at not finding faces unattached to heads is the ironic reason why AI is so unreliable.
Sage advice to would-be AI developers: Invent an algorithm that can see the Virgin Mary on a grilled cheese sandwich and you may have something worthwhile.
chirho said...
This story will confirm a belief that racism is embedded in every part of society, even in the 1s and 0s of computer codes.
WRONG! Critical Digit Theory tells us, that ONLY ONES are racist.
As we ALL know; Racism is Prejuduce plus power, right?
So, only digits that carry a value, are Capible of having power; right?
And in a binary system, that means Only a One
A zero, represents The Lack of power (right?); it is Merely a placeholder. It is the ones running the show
So, until we GET RID of the ones, Racism will remain inherit in our number systems
Some of you, may think, that i am advocating the Removal of the Binary system
While that IS TRUE, let's not forget, That ANY digit carrying a value, is INHERITLY RACIST
Until we force math, to ONLY USE ZEROS.. MATH is RACIST
2+2=4? you can NOT BE more racist than that!!!
0+0=0 THIS IS THE ONLY ACCEPTABLE ANSWER... Until we are ALL zeros, NONE of us will be free
Blast from the past
Flip Wilson Ugly Baby Routinr
@Questor - Yes, we are definitely in the "parvorder Catarrhini, which includes the ld World monkeys and humans, but not the New World varieties." We share a common ancestor with old world monkeys. But most taxonomists think of the old world monkeys as a subset of that clade, and don't refer to the gibbons, siamangs, orangs, gorillas, chimps, or humans as monkeys.
For those who care, a nice cladogram occurs about two thirds down this.
Maybe they could strike the 'primates' and substitute 'homonids'
didn't see the images, but I was just listening to scott adams, and he brought this up, pointing out that the cops were possibly the trigger for aI because their riot gear caused them to be uniformly bulky & black clad.
interesting.
My son works in AI, and they have found that they have a problem that AIs are really racist. If you give them a bunch of potential employees and their info, it doesn't matter if you remove the race: the AI figures it out from other information and scratches those candidates because they are more likely to have criminal backgrounds or bad work ethic or who knows what else.
Quaestor @1549: Brilliant comment, thanks.
An AI is neither sentient or sapient. It can be inaccurate but not "racist".
The real problem is when it provides conclusions that contradict the leftist worldview. It's even worse if those conclusions are accurate.
Lars Porsena writes, "Maybe they could strike the 'primates' and substitute 'hominids'."
You won't find the term hominid, (a member of the family Hominidae) in most work published in the last thirty years because it is so broad. Back when Johanson et al. published the "Lucy" find Homo seemed to be a remarkably un-bushy clade for such a widespread (East Africa to Indonesia) genus. The hard work seemed to be connecting A. afarensis to the Hominidae. Since then, with the description of many new Homo species with very murky relationships to ourselves, such as H. ergaster, H. denisova, H. floresiensis, and others not yet formally described, much of the discussion today, particularly in the popular press, revolves around a recently coined term, hominin, a member of the tribe* Hominini.**
This has confused many laymen, leading many to believe hominid to be an obsolete term. Not true, it's just that the fashionable work is being done closer to home, as it were.
* Tribe is also a recently invented (and controversial) taxonomic ranking, however, without it some evolutionary relationships are just too lumpy to be intelligible. One suspects the taxonomists will need to add several more rankings under tribe before things are clear. (Clique? Gang? Circle?)
** Hominini includes Homo and its immediate bipedal ancestor(s), and therein lies the rub. It's not clear that bipedality arose only once and that bipedal species was ancestral to every member of our genus. It is possible that about four and a half million ago two or more closely related anthropoid species evolved bipedality in response to similar conditions prevailing over a wide area of Africa, and that two or more of these two-footed apes gave rise to manlike descendants, which would mean that Homo might be a paraphyletic taxon! The problem seems to emanate from the dominance of Louis S. B. Leakey, who was obsessed with skulls and brains to the point that other hominin fossils, particularly toe bones, knee bones, and hips may have been overlooked. (Leakey dominated post-war paleoanthropology mainly due to funding and publicity provided by the NatGeo organization.) It may well turn out that the real story of human evolution is a "below the waist" matter.
It may well turn out that the real story of human evolution is a "below the waist" matter.
So much of evolution is :)
Has anyone ever seen Elijah Cummings and Harambe together in the same room?
Tales of the Interregnum. In another fifty years, what the algorithms and our robot overlords do will go unchallenged.
Models predicting climate change use the same technology.
Primate boy: Do not try and identify the primate. That's impossible. Instead... only try to realize the truth.
Neo: What truth?
Primate boy: There is no primate.
Neo: There is no primate?
Primate boy: Then you'll see, that it is not the primate that identifies, it is only yourself.
If the Facebook prompt had asked "would you like to watch more videos like this?" noone would have complained. Then the AI could have gone searching for more gorilla movies without anyone the wiser.
AI cannot distinguish a human being from an ape or a lemur, but it is totally fine for imposing a political censorship regime.
One wonders what Apple will turn up when it starts scanning everybody's iPhone for child porn, with no probable cause or even any cause whatsoever.
Post a Comment