Former author of one of the top 5 facial recognition servers in the world for multiple years running, here's what's going on: the industry has solved this issue, but the potential clients are seeking the lowest bidder, and picking the newer companies, the nepostically created not really players but well connected, and those companies have terrible implementations. This is not a case of the technology not there yet, we solved all these racial bias issues 10 years ago. But new companies with new training sets and new ML engineers that do not know any of the industry's history are now landing contracts with terrible quality models, but well connected sales channels.
This study finds a higher rate of correct identification for black people than for other ethnic groups, whereas a few years ago the problem seemed to be that the software was less effective at identifying black people.
Do you have some insight about why this reversal might have occurred?
To have a high quality facial recognition system it needs to include every possible combination of ethnicity, in addition to all of those they each need to include variations of daylight, of dappled light, of partial obscuring, night time illumination, across every variation of season, variations of expression and face angle, across variations of weather, variations of distance, across variations of things placed on a person's face, and then across all kinds of variations of video compression. All these face image variations in the training set enable the trained model to find and track the features that persist through all these variations. In truth it requires hundreds of millions of facial images to create an accurate facial recognition system. Most new companies and many that have been around for respectable periods are not realizing how much data collection, annotation and additional variation creation it requires for a high quality FR training set. The company I worked at spent 20 years collecting laser scans of real people to then create the augmented real person data set with several hundred million faces.
Your problem is evident in your question. What does “black” mean? It’s entirely subjective. Dwayne Johnson? Liv Tyler? Nelson Mandela? Barack Obama? Mariah Carey?
This is a semantic issue. Ethnic groups are constructs. A system which misidentifies people identifies all people poorly.
It doesn’t track across regions either. People labelled “white” by law in some countries (Brazil, South Africa, etc) would be classified as “black” elsewhere.
In England, the example here, we do not classify people the way the Us does, with its history of “one drop” politics. Many British people considered “white” are “black” in the US.
There is no scientifically valid way of defining who counts as “black” so any discussion of tuning a system based on this definition is a disaster.
Even the people commenting are talking about different groups based on their own culture and prejudices.
I believe the term "Black" in reference to a person when discussing the topic of facial recognition is only used in journalism. There is no "Black" in the facial recognition industry. There really is no identification of ethnicity in facial recognition. It is all just variations of human appearance, in a unbroken spectrum. The natural and ever present population of mixed race people basically destroy any sense of "race" or "ethnicity" within the software. The ONLY time race and ethnicity are included in facial recognition discussions is when some group trains an algorithm with biased data, creating a biased trained algorithm. That is a human failure to understand the problem they trained their data, not grasping the lack of critical data and its impact on the trained model's use. The technology itself operated exactly as designed, it was literally humans not understanding the subtle nature of what they were doing that is the issue.
> Many British people considered “white” are “black” in the US.
I'm also British, can you give an example of that? A minor celebrity/TV personality say?
To the extent you have a point though I think it's irrelevant anyway—they paused the programme because they found, according to whatever definition they measured with, it had that skew.
How recently? We had a home security camera and every time our (Black) son walked up to the door, the camera would classify him as an “animal”. This was as recently as 2022
In the other direction, my camera regularly identifies cats, crows, and shadows as people. I think recognition in security cameras has a very long way to go.
This is actually more (socially/ethically/philosophically) interesting than one might assume from the headline: it's not false positives, it's that it's more effective (correctly identifies someone is on a watch-list) for one group than another within a protected characteristic.
So essentially they're pausing the use of it because it works too well for group A / not well enough for group B, potentially leading to disproportionate (albeit correct) arrests of group A.
Absolutly impossible to condone further structural bias against a minority, and just ignore the free "white pass" built into the software, and esspecialy troubling that it passes white women, the most.
The only possible action is to reject and dissable any system with a racial bias, investigate how such a thing happened, with a very pointy look for intent on the part of the vendors, who would then qualify for bieng housed in one of his majestys facilities for persons such as these.
If you start with hypothetical demographic groups A and B that are for all intents and purposes exactly identical, but you implement a system such that if A commits a crime they have a 10% chance of being caught and if B commits a crime they have a 50% chance of being caught, you will achieve the following:
1. More short-term crime prevention than a system catching 10% of A's crimes and 10% of B's crimes (good!)
2. Enforce a societal belief that A is intrinsically better than B (bad!)
3. Disproportionately burden children, families, and communities in B than A, causing them to indeed perform worse in everything than those in A (bad!)
4. As a result of 2 & 3 it is not a stretch to say simply causing B to do more actual crime (potentially negating point 1 entirely)
If you believe that crime enforcement is not for the sake of vengeance but instead something done to improve the well-being, safety, and happiness of citizens, you may see that inequality=bad just as crime=bad. How to best solve this trolley problem is complicated but it's important that people are aware that it is complicated before firing off an answer.
it is FALSELY unidentifying people, which makes the harware, software, sales, implimentation of the whole system a criminal enterprise, which it is. Kudos to the police for rejecting this racist biggoted unjust criminal software implimentation.
See, what you've said is precisely "structural bias against a minority", or "systemic injustice". Then again, the elites are, technically, also a minority as well, and we all know how well letting their crimes slide works out for the rest of the society.
> the system was more likely to correctly identify men than women and it was “statistically significantly more likely to correctly identify black participants than participants from other ethnic groups”
Technology has moved on a lot no doubt, however, studies were finding the opposite (and with order of magnitude errors) as recently as 2020 with a lazy google literature search
> these algorithms were found to be between 10 and 100 times more likely to misidentify a Black or East Asian face than a white face
> the system was more likely to correctly identify men than women and it was “statistically significantly more likely to correctly identify black participants than participants from other ethnic groups”.
I am genuinely unsure what's going on.
My understanding of the article is that the system is problematic because it is more likely to correctly identify black people than "other ethnic groups". Is that right?
“statistically significantly more likely to correctly identify black participants than participants from other ethnic groups”.
Great. Wasn’t the problem before always that it couldn’t correctly identify non-white people?
It does it accurately now. That is somehow also a problem? It should make more mistakes?
47 comments
Do you have some insight about why this reversal might have occurred?
This is a semantic issue. Ethnic groups are constructs. A system which misidentifies people identifies all people poorly.
It doesn’t track across regions either. People labelled “white” by law in some countries (Brazil, South Africa, etc) would be classified as “black” elsewhere.
In England, the example here, we do not classify people the way the Us does, with its history of “one drop” politics. Many British people considered “white” are “black” in the US.
There is no scientifically valid way of defining who counts as “black” so any discussion of tuning a system based on this definition is a disaster.
Even the people commenting are talking about different groups based on their own culture and prejudices.
> Many British people considered “white” are “black” in the US.
I'm also British, can you give an example of that? A minor celebrity/TV personality say?
To the extent you have a point though I think it's irrelevant anyway—they paused the programme because they found, according to whatever definition they measured with, it had that skew.
Frankly, I'm skeptical, but I'm willing to be convinced by reputable evidence.
So essentially they're pausing the use of it because it works too well for group A / not well enough for group B, potentially leading to disproportionate (albeit correct) arrests of group A.
1. More short-term crime prevention than a system catching 10% of A's crimes and 10% of B's crimes (good!)
2. Enforce a societal belief that A is intrinsically better than B (bad!)
3. Disproportionately burden children, families, and communities in B than A, causing them to indeed perform worse in everything than those in A (bad!)
4. As a result of 2 & 3 it is not a stretch to say simply causing B to do more actual crime (potentially negating point 1 entirely)
If you believe that crime enforcement is not for the sake of vengeance but instead something done to improve the well-being, safety, and happiness of citizens, you may see that inequality=bad just as crime=bad. How to best solve this trolley problem is complicated but it's important that people are aware that it is complicated before firing off an answer.
Selective enforcement has been used historically to persecute certain minorities and that's not acceptable in a functioning society.
> the system was more likely to correctly identify men than women and it was “statistically significantly more likely to correctly identify black participants than participants from other ethnic groups”
Technology has moved on a lot no doubt, however, studies were finding the opposite (and with order of magnitude errors) as recently as 2020 with a lazy google literature search
> these algorithms were found to be between 10 and 100 times more likely to misidentify a Black or East Asian face than a white face
https://jolt.law.harvard.edu/digest/why-racial-bias-is-preva...
> more likely to correctly identify men than women.
> more likely to correctly identify black participants than participants from other ethnic groups.
> AI surveillance that is experimental, untested, inaccurate or potentially biased has no place on our streets.
I wonder if they're more worried about putting too many men in prison or too many black people.
> the system was more likely to correctly identify men than women and it was “statistically significantly more likely to correctly identify black participants than participants from other ethnic groups”.
I am genuinely unsure what's going on.
My understanding of the article is that the system is problematic because it is more likely to correctly identify black people than "other ethnic groups". Is that right?
Great. Wasn’t the problem before always that it couldn’t correctly identify non-white people? It does it accurately now. That is somehow also a problem? It should make more mistakes?
Essex police, well aware of all the issues before using it, pause use until expected bad publicity dies down
Or
Essex police chosen as force to take some flack for the issues while other forces steam ahead