Slashdot Mirror
Facial Recognition Is Accurate, if You're a White Guy (nytimes.com)
Facial recognition technology is improving by leaps and bounds. Some commercial software can now tell the gender of a person in a photograph. When the person in the photo is a white man, the software is right 99 percent of the time. But the darker the skin, the more errors arise -- up to nearly 35 percent for images of darker skinned women, the New York Times reported, citing a new study. From the report: These disparate results, calculated by Joy Buolamwini, a researcher at the M.I.T. Media Lab, show how some of the biases in the real world can seep into artificial intelligence, the computer systems that inform facial recognition. In modern artificial intelligence, data rules. A.I. software is only as smart as the data used to train it. If there are many more white men than black women in the system, it will be worse at identifying the black women. One widely used facial-recognition data set was estimated to be more than 75 percent male and more than 80 percent white, according to another research study.
White people's faces reflecting more light is problematic.
Darker colors provide less contrast. Less contrast means features are more difficult to make out.
Combine that with the typically horrendous lighting video cams face and you have a situation where recognition fails.
When Fascism comes to America, it will call itself Anti-Fascism, and tell you to give up your guns.
Always knew that machines were bigots.
Have you ever watched some war move where all the actors are relatively unknown and are sporting a buzz cut?
Can't tell one white guy from another, especially if they have similar builds.
And when they later use camo face paint...forget it. All you can rely on is their voices
When Fascism comes to America, it will call itself Anti-Fascism, and tell you to give up your guns.
It's official, I can no longer differentiate between alt-righters and SJWs, both just say "we're the victim and everybody else are raping/genociding/whatever us"
Avantgarde Hebrew science fiction
What's the difference between porn and erotica?
Lighting.
--
BMO
This reminds me of that ridiculous article (and accompanying video) saying that color film was biased towards white people. Around 3:30 in the video they have white and black people stand in front of a face-following camera and it doesn't work for the black people. Everyone acts like this is some sort of Harry Potter wizardry against the black man keeping him down when it's vastly simpler than that.
For progressively darker skin, progressively higher light on that skin is required to reveal its contours. The fundamental problem is that white and light-skinned brown people have their normal skin color shades in the midtones when a scene is properly exposed while darker-skinned brown and black people are closer to shadows. To expose properly for facial recognition of dark brown or black skin, you have to overexpose the midtones to bring up the shadows. Since people rarely take photos on purpose that are exposed for the shadows while blowing everything else out, it should be fairly obvious that facial recognition (and early ISO 32 color film and small-sensor cameras like webcams and phone cameras) will have a very hard time with dark skin. Sure, it could be a lack of data in some instances, but it's far more likely to be the fact that the skin absorbs more light and photographs are generally exposed too low to reveal enough detail for the machines to analyze.
If you think this is "racist" you're saying that the nature of light itself is racist. I don't feel like I should have to explain why that position is really stupid.
Most of the time these cameras would work just fine if they adjusted the exposure properly. They have auto exposure but it's tuned either for general photography or white skin.
The fix isn't that complex, and some kind of calibration could be done when setting up face recognition.
const int one = 65536; (Silvermoon, Texture.cs)
SJW, n: "Someone I don't like, and by the way I'm a fuckwit" - AC
Then there is no discrimination. Kinect infrared for example does a great job leveling the playground.
I don't really think that is true, but it is true you need different lighting to pick up the nuances in a black face. Movies and TV shows for years have lit black actors in ways that wash out their faces by using rules of thumb that work for white faces. It's only in the last few years that we're seeing black actors properly lit.
Post may contain irony: discontinue use if experiencing mood swings, nausea or elevated blood pressure.
Also known as the base rate fallacy. If you're looking for a needle in a haystack, an algorithm which correctly distinguishes them 99% of the time is useless.
Post may contain irony: discontinue use if experiencing mood swings, nausea or elevated blood pressure.
Racist? Not really - just that one learns to identify people by looking at people. So Africans generally think Asians and "whites"* look all the same, Asians generally think "whites" and Africans look the same etc.
Nothing strange or racist about that.
Then add the fact that many people have problem identifying others from facial features alone we get a cultural aspect of this "look alike" thing.
The tabby takes great pictures
What sort of camera have you trained that cat to use when he's taking those pictures?
Don't disappoint your bird dog. Go to the range.
Actually that is literally the definition of racism.
I just looked it up and it's literally not the definition of racism.
Trump's a white guy isn't he?
Orange is the new white?
No, it will find "needles" all over the place, laced through the haystack. Just one of those tens of thousands of "needles" will actually be a needle.
Post may contain irony: discontinue use if experiencing mood swings, nausea or elevated blood pressure.
That's beside the point in this context, because with most facial recognition cameras, you don't control the light. Most of the time, it's going to be daylight, and sometimes street lights or store lights.
That darker objects reflect fewer photons is always going to be the case. That's not racial bias.
Fitness trackers with optical heart readers have a similar problem, where the light does not penetrate as well for darker skin. Some auto-adjust, with the unfortunate side effect of shorter battery life for dark skinned people. And some blast at too high intensity for everyone, and cause burn marks over time for the fairest skinned people, and make their heart beat harder to read.
There's no real solution that fits everyone. People are different, and "one size fits most" is not going to be the optimum solution for either facial recognition, iris scanners, optical pulse monitors, or anything else. Accepting that someone from Iceland is likely going to absorb less photons than someone from South Sudan, and thus optical based equipment will work differently isn't racial injustice.
Someone needs to test whether humans, also, decline in speed or accuracy of facial recognition when dealing with darker shades of skin colour.
I know for certain that I have more trouble reading facial emotion from black people than white people. The naive response is that I live in a city that's 95% white. But I've been able to convince myself that this is the correct explanation. I simply feel like I have less visual data than I would otherwise at the same point in the cognitive process.
Suppose I lived in a troop deployment in Afghanistan, and 90% of the people around me wore camo all the time. Would I actually become better at recognizing camo than civilian gear? But this is, indeed, the converse implication of the naive hypothesis.
There are populations in Brazil that experience the entire range of skin tones on a daily basis. These populations could be tested for recognition rate/accuracy for lighter and darker test cases.
I highly suspect that darker skin tone has a detectable coefficient of identity camouflage, also in human cognition.
In our increasingly Orwellian society, I would be quite happy to have facial recognition technology be less effective on my skin tone (fair).
A closed mouth gathers no foot.
Facial morphology refers to the various traits and features in a face. For example, the distance between the eyes, or the eye slant, or cheek gaunt or whatever.
'White' people have the broadest range of diversity, in part because aside from the skin color, there's a lot of differences. Certain Asians, like the Han Chinese, have some of the least diversity (google for iphone face recognition matching two Chinese co-workers).
If you pick 20 key features as your unique code, and each of those key features has 20-30 distinct possible values, you can rely on reasonable uniqueness, even when some of those values have inter-relationships. When the diversity goes down, and 10 out of the 20 are not unique, and when the range of values those have is between 3 and 5, well, you'll have a lot more trouble differentiating people.
In fact, a studies shows that among a given ethnic group, actual real life people perform facial recognition on only a few features, but those features are always those traits that show the most variation. When you apply that same algorithm to another ethnicity, it doesn't work so well. You get racist-seeming phrases like, "They all look alike to me," when really the issue is that your specialized detection algorithm was never meant to deal with their differences. ... and every group has this blindness. The one thing that's amusing is that because whites tend to have a large variety, they're the easiest to uniquely identify regardless of your personal/cultural/ethic technique. So, you can say things like "I can tell all you white people apart, you're racist for not being able to identify ME!" and think you're on the moral and ethical high road, when in fact, the situation is different from the other side.
Actually it is on point with regard to testing.
Anyhow, what you're saying black faces have less luminance -- well sure. That says nothing about contrast. The idea that there's less detail there to be seen is a result of looking at poorly lit photographs.
The upshot is that if the problem is in the sensor's ability to handle a certain luminance range, by that theory the algorithms should perform better on black faces in very brightly lit conditions.
Post may contain irony: discontinue use if experiencing mood swings, nausea or elevated blood pressure.
There's more than enough contrast for a convolutional neural net to work with. You could probably turn it all into 4 bit grayscale before training and still get excellent results.
The explanation isn't what you propose, but unbalanced training sets. Failing any face is equally bad for the training algorithm. Whether its errors are equally divided among all subgroups, or concentrated in one of them, is equally good for the algorithm. Since it has more data on whites, it can profit more from focusing on features characteristic of them.
You can change the training to penalize having a high error rate for subgroups. But this comes with tradeoffs. Better is to get more training data from the difficult subgroups, and train a better algorithm overall. The best way to get an algorithm that makes few errors on blacks (or any other subgroup) may just be to get an algorithm that makes few errors, period.
Actually, light skinned people have more variations in facial shape than do other groups. Second is darker skinned people. Orientals have the fewest.
OTOH, in any particular area, the people from outside the area are likely to have more variation, because they have a wider variety of ancestors.
That said, this *is* a bit strange, because the greatest genetic variation is among the population native to Africa. (Note I'm not even including the Australian aborigines. Which are a part of the facial variation of the darker skinned people.) So one would expect the largest variation among the darker skinned people.
Additionally the homogeneity of the orientals is probably due to their long period of civilization. This is a guess, but it's a reasonable one. So there was a longer period of undisturbed gene flow among groups. Even so there are distinct sub-groups, just not as many as among other categories.
A problem with this analysis is that I didn't include the population of the Indian sub-continent, as I couldn't figure out in which group to place them. They have darker skin colors, but have facial features that more closely align with the lighter skinned peoples. This is readily explained by historical analysis, but it does make categorization difficult.
I think we've pushed this "anyone can grow up to be president" thing too far.
The Summary said it was because they were underrepresented in the training data set.
That's what you should first assume when an AI system fails at some particular kind of categorization, so it should hardly be surprising.
I think we've pushed this "anyone can grow up to be president" thing too far.
Yes, a story listing examples of how white people can point guns at cops and live and black people cannot.
lol. If you're stupid enough to think that a cop with a gun pointed at him gives a shit about the colour of the hand holding it, you can't possibly expect to be taken seriously in these kinds of discussions.
Of course he gives a shit. There are centuries of examples in the United States to police giving white people the benefit of the doubt when black people would have been put under the jailhouse.
I can't believe you're stupid enough to think that kind of long-standing prejudice just suddenly went away in the past 20 years.
You are welcome on my lawn.
Sorry, I misunderstood your previous statement. What you just said makes sense. If I understand what you are saying, attempting to compensate by overexposing will cause the face to look unnatural.
You're right, they shouldn't overexpose as a solution since the point was to identify people from a normal picture. As opposed to using over exposed pictures that are focusing on image recognition. Unless it was an HDR+ picture, it would look really bad. Even an HDR+ picture would have issues unless people wanted portraits of themselves with unnatural skin tones.
Why are we calling this bias? White males have the most range of unique identifying characteristics:
* Beards
* Moustaches
* More tonal contrast
* Difference in eye and hair colour
Strange you would make this about social justice, and not just an interesting technology problem. TFA doesn't mention racism or anything... It's almost like you are some kind of social justice obsessed warrior who has to bring it into every conversation.
Have to agree with you about the sock puppets though.
const int one = 65536; (Silvermoon, Texture.cs)
SJW, n: "Someone I don't like, and by the way I'm a fuckwit" - AC