New research shows driverless car software is significantly more accurate with adults and light skinned people than children and dark-skinned people.
New research shows driverless car software is significantly more accurate with adults and light skinned people than children and dark-skinned people.
WEIRD bias strikes again.
deleted by creator
A stealth bomber gives less signal because of angles and materials and how they interact with radar, not because they are small or painted a dark color.
If a dark skinned person and a white skinned person are both wearing the same pants and long sleeved shirts, why would skin color be a factor beyond some kind of poorly implemented face recognition software like auto focus on cameras that also don’t work well for dark skinned folks? Especially when some of the object recognition is just looking for things in the way, not necessarily people.
No, it is not some simple explanation based on people’s eyes from the driver’s seat while driving in the dark. It is a result of the systems being trained based on white adults (probably men based on most medical and tech trials) instead of being trained on a comprehensive data set that represents the actual population.
deleted by creator
While some of these cars use radar to an extent. I believe this is mostly focusing on image recognition, which is from a camera. Both are distinctly different in how they recognize objects.
Image recognition relies on cameras which relies on contrast. All of which is dependent on light levels. One thing to note about contrast is that it’s relative to its surroundings. I think this situation is more similar to your eyes recognizing things while driving in the dark than you think. I suggest you research how these things work before making claims.
So white people have higher contrast than dark skinned people?
Yes, in fact. This has been a huge challenge in photography algorithms for decades.
HP cameras couldn’t detect black people in 2009: http://edition.cnn.com/2009/TECH/12/22/hp.webcams/index.html
Google classified black people as gorillas in 2015: https://www.theverge.com/2018/1/12/16882408/google-racist-gorillas-photo-recognition-algorithm-ai
Zoom had issues with black faces and dark backgrounds in 2020: https://onezero.medium.com/zooms-virtual-background-feature-isn-t-built-for-black-faces-e0a97b591955
A quick primer in colour: recall that light colours reflect more light than dark colours. This means image recognition devices relying on cameras using standard spectrums (i.e. not infrared) receive less light into the sensor when pointed at someone with dark skin. The problem is constant, but less pronounced depending on the background. That is, a black person against a white background would be easier for an algorithm to identify as a person than said black person against a mixed or dark background.
All of those had issues for the sensors and recognition aoftware because their data set to determine what a face is was mostly white people.
Just because something is harder doesn’t excuse then for not putting in the effort to get it right.
It’s not necessarily effort. Data can be expensive and difficult to obtain. If the data doesn’t exist then they have to gather it themselves which is even more expensive.
I agree that they should be making sure they can account for both cases as much as possible. But you have to remember that from the frame of reference of the model being trained and used in these instances, the only data they’re aware of is the data they were trained on and the data they are currently seeing. If most of the data samples in the entire world feature white people 60% of the time it’s going to be much better at recognizing white people. I don’t think anyone is purposely choosing to focus on white people; I think that those tend to be the data samples that are most easily obtained or simply the most prolific.
I also think we need to take into account quality of data. As mentioned before, contrast plays a big role in image recognition. High contrast with background results in, on average, better data samples and a better chance of usable data. Training models on data that is not conclusive on ambiguous can lead to ineffective learning and bad predictive scores.
I don’t think anyone is saying this isn’t a problem but I also don’t believe that this is a willful failure. I think that good data can be difficult to get and that data featuring white people tends to have easier time using image recognition successfully.
Someone else mentioned infrared imaging, which is a good idea but also more money and adds an extra point of failure. There are pros and cons to every approach and strategy.
Cost being used as an excuse not to expand the data set to represent all types of people is just excusing systemic racism and other discrimination. For example, if the system requires two arms for it to recognize a person that is also a problem, because a person comes in a wife variety of shapes, sizes, and colors.
If the system can’t handle that then it doesn’t regocnize people. If it costs too much to do right, then that means they can’t afford to do it at all.
In some cases the data sets were only white, but engineers have been cognisant of this issue for decades so I don’t think that’s as common as you might believe. More frequently it’s just physics.
As for “putting in the effort,” companies are doing this, to their detriment. Ensuring that a small proportion of their customer base has a perfect experience is very expensive. In business the calculation between cost and profit is very important. If you’re arguing that companies should provide unprofitable products so that your sensibilities can be assuaged then I disagree. No company has a duty to provide a product to you.
A company making driverless cars damn well does have a duty to make sure their program doesn’t run over children.
We are talking about that portion of the population being hit by cars.
It’s both. The system is racist because of how it was trained and because its developers were not black, therefore “it worked for them” during development. And because black people are harder for cameras to see, especially in low light environments.
Even with clothes on, the dark skin, in a dark environment, “breaks” the “this is human” pattern that the ai expects to see, since the ai can see only the clothes. It is like camouflage. Can the ai “see” a pair of pants? Maybe, eventually but it still reduces the certainty, since the ai sees fewer “signs”.
Cameras should be using infrared to look for objects in the dark and not fucking hoping it looks slightly less dark than the surrounding pixels. It being “dark” is not an excuse. Cars drive at night and need to be engineered around that fact.
Edit: note this is about cameras. Ideally, you’d use radar which wouldn’t care but if you are just dual purposing cameras used for driving, this is the bare minimum.
deleted by creator
These systems are often trained on data obtained from driving the car around. I think the only real solution would be planning routes through more diverse neighborhoods. Although any company that is taking this seriously from a safety perspective has multiple radars and a top mounted LiDAR on their vehicles. Those sensors should be sufficient for detecting humans regardless of race even in a completely dark environment. Relying solely on camera data is just asking for problems for this and many other reasons.
Except that’s not the source of this problem. AI can be great at detecting patterns with little data, if it’s properly trained. But this article is clear that the reason of this failure is in the lack of training data. This means that the AI never learned kids and dark-skinned people exist and it’s unreliable in detecting them.
It is easier to see dark object on bright background.
That’s only part of it though. This issue is almost as old as we have had similar image/facial recognition technologies. Data is where models get their conclusions from.
Yes but isn’t it easier to say RACISM
You don’t know that.
Speaking as someone who inherited a computer vision codebase from Asian devs and quickly found that it didn’t work on white skin…
Implementation decisions matter, and those decisions will always be biased towards demonstrating successful output for the people who plan, bankroll, and labor on the project.
How much of the 20% or 7.5% difference in detection is due purely to inevitable drawbacks of size and skin tone?
Who knows.
What we do know is that we did measure a difference, and we do live in a culture where we’re more likely to hear a CEO say:
“It works!” …and then see an article months later that adds “…except for children and black people.”
rather than:
“It doesn’t work!” …and then see an article months later that adds “…except for adults and white people.”
And that fact means we should seriously consider whether our attention (and intention) is being fairly applied here.
Wow. that’s all kinds of incorrect
It is absolutely data training bias. Whether it is the data that ML was trained on or the data that programmers were trained on is irrelevant. This is a problem of the computer not recognizing that a human is a human
It is not. See below:
No, not if the scale of your hardware is correct. A 3’ tall human may be smaller than a 6’ one, but it is larger than a 10” traffic light lens or a 30” stop sign. The systems do not have quite as much trouble recognizing those smaller objects. This is a problem of the computer not recognizing that the human is a human.
Also no. If that were the case, then we would have problems with collision bias against darker vehicles, or not being able to recognize the black asphalt of the road. Optical systems do not rely on the absolute signal strength of an object. they rely on contrast. A darker skin tone would only have low contrast against a background with a similar shade, and that doesn’t even account for clothing which usually covers most of a persons body. Again, this is a problem of the computer not recognizing that the human is a human.
No, they have different signals. that signal needs to be compared to the background to determine whether it exists and where it is, and then compared to the dataset to determine what it is. This is still a problem of the computer not recognizing that the human is a human.
Close, but not quite.
This is a problem of the computer not recognizing that the human is a human.
You sound like an imaging specialist with no experience
I’m sure that will be of great comfort to any dark-skinned person or child that gets hit.
If those are known, expected issues? Then they had better program around it before putting driverless cars out on the road where dark-skinned people and children are not theoreticals but realities.
In order to make the software detect the same you have to make it detect white adult less.
Comparing the performance between races says nothing about how safe a driverless car is. I am sure that the chances of a human hitting a dark skinned person dwarfs the chances of a driverless car. Trying to convince people driverless cars are racist only delays development, adoption and lawmaking which means more flawed meatbags behind the wheel which means more car accident deaths.
What he’s saying is these aren’t issues, they’re like saying a masculine voice can be heard from further away. Deeper voices just carry better
Part of it is bias/training data - we can fix that. But then you’re still left with the fact children are smaller and dark skinned people are darker - if you use the human visible range of light (which most cameras do), they’re always going to be harder to detect than larger more reflective people.
Our eyes and brains have an insane ability to focus and deal with varying levels of light, literally each cell adapts individually to each wavelength. We don’t have much issue picking out anyone until it becomes extremely dark or extremely far away - it’s not because the problem is easy, it’s because humans are incredible at it
Thank you.
You seem to be one of the people who understand this better.
And even humans are not incredible at it. It’s just inherently harder to identify the areas where there are less signal. I’d love to see a study, but see my edit and actually quantifying the equality we’re after.
Reality/physics/science/PDEs (whatever) work on “differences”. The less difference, the harder.