Slashdot Mirror
Google Executive Warns of Face ID Bias (bbc.com)
Facial recognition technology does not yet have "the diversity it needs" and has "inherent biases," a top Google executive has warned. From a report: The remarks, from the firm's director of cloud computing, Diane Greene, came after rival Amazon's software wrongly identified 28 members of Congress, disproportionately people of colour, as police suspects. Google, which has not opened its facial recognition technology to public use, was working on gathering vast sums of data to improve reliability, Ms Greene said. However, she refused to discuss the company's controversial work with the military. "Bad things happen when I talk about Maven," Ms Greene said, referring to a soon-to-be abandoned project with the US military to develop artificial intelligence technology for drones. After considerable employee pressure, including resignations, Google said it would not renew its contract with the Pentagon after it lapses some time in 2019. The firm has not commented on the deal since, only to release a set of "AI principles" that stated it would not use artificial intelligence or machine learning to create weapons.
The technology behind FaceID has no bias. It works really well - if given the right training data. Now it could easily be that the training data you are feeding it is biased in some way, but that is why extensive testing of the resulting recognition engine you have built is key, so you can go back and correct training data...
Because training neural networks is kind of a blackbox, it's sometimes hard to say what kind of bias you may have built in. the Amazon system recognizing a set of politicians as criminal might be down to the lighting used in the picture being a lot like mug shot lighting!
Or who knows, maybe it's latched onto specific micro-expressions of criminals and the politicians it identified really are criminals, we just don't know it yet... :-)
"There is more worth loving than we have strength to love." - Brian Jay Stanley
That's really broken. It missed at least 500 more...
The problem is the amount of light the camera sensors receive. Darker faces reflect less light, and thus the camera sensor gets less data to work with making algorithms based on that data less accurate at identifying darker faces.
This presents an obvious solution. To further the goal of eliminating racial bias, we need to turn off all the lights. That means all light bulbs need to be banned, and existing ones destroyed. NASA should launch a huge unfurling disk to block out the sun and leave the planet in perpetual darkness. Newborns should have their eyes surgically removed upon birth (they won't suffer because they won't know what they're missing). Only then can we be free of the evil racial bias being promulgated by light.
The technology reflects the biases of its inventors
The "technology" behind this is neural networks. How do they reflect bias at al?
They are nothing more than the ultimately transparent black box, reflecting whatever you choose to put into it...
They are so non-biases in fact, the same technology is used to detect if something is a cat or a slice of pizza, or even if tumors are cancerous or not. Yet you could claim racial bias (you did not state that but it was implied).
Like I said, care needs to be taken both in training and in testing, that is where bias may be introduced. But the technology itself is inherently unbiased and a great tool, if used correctly.
"There is more worth loving than we have strength to love." - Brian Jay Stanley
The technology behind FaceID has no bias. It works really well - if given the right training data.
Given that clearly there isn't a set of "right training data" available that sounds like a hasty conclusion without evidence. Your argument is circular. You say the technology has no bias but proving that it has no bias requires feeding it an unbiased data set which hasn't happened. So neither of us knows if there is an inherent bias built into the system or not. Maybe there is and maybe there isn't but you don't have the data to say either way.
Furthermore it's more complicated than just the training data set. There is code that determines the logic used to analyze the data and that code is subject to biases from the person programming it. They do not have to be conscious biases either. That's not to say it is an irreducible problem but it has to be considered.
Now it could easily be that the training data you are feeding it is biased in some way, but that is why extensive testing of the resulting recognition engine you have built is key, so you can go back and correct training data...
Who is watching the watchmen? The ones checking the data, providing the data, and analyzing the data, and revising the data are people and people have biases. Guarding against and when necessary removing those biases can prove rather tricky especially when the training data sets used are by and large positively loaded with biased data.
The "technology" behind this is neural networks. How do they reflect bias at al?
Several ways but the two most straightforward are Bad training data and Bugs. Bugs are an issue in ANY software and neural networks are no different and bugs can result in biases. And of course train it with bad data and you'll get biased results. We've seen examples of both cases.
the two most straightforward are Bad training data
Which is not inherent in the technology as I've pointed out in every post.
and Bugs. Bugs are an issue in ANY software and neural networks are no different and bugs can result in biases.
That reflects a lack of understanding of how neural networks work. They actually are not buggy themselves as they are very simple; all bugs are entirely represented in training data and testing, not in the software itself. You put something in the black box and it classifies it in some desired way, the way the box works is all about the training data.
Furthermore it's a bit odd to claim bugs are a kind of "bias" as usually they are more about failure than bias...
"There is more worth loving than we have strength to love." - Brian Jay Stanley
It's harder to see the contours of a dark-colored shape (ie a face) than a white one.
Seriously, people, how are we going to get around that?
-Styopa
Given that clearly there isn't a set of "right training data" available>
Apple got this right, but you are right that generally there is not a "right training set" for facial recognition.
There cannot be though because it all depends on what you are trying to do. What is right for one purpose would be wrong for another.
Your argument is circular. You say the technology has no bias but proving that it has no bias requires feeding it an unbiased data set
That reflects a total lack of understanding of neural networks and facial recognition. It provably has no bias because it can and is applied to anything, it's simply to general in nature to have any bias. It's like claiming electrons have a bias as to the humans they prefer... it makes no sense.
"There is more worth loving than we have strength to love." - Brian Jay Stanley
They are reporting this as if the false positives are either matches or not match. In reality, there are associated levels of confidence, and a match is likely anything more than 95% confidence. Less light means there is more similar data in a photo. What they really need to do is run a different confidence level on faces that reflect less light. Maybe even run a completely separate facial recognition algorithm so the accuracy of the better data is not muddying the confidence levels of the worse data.