Once again, racial biases show up in AI image databases, this time turning Barack Obama white
- Reference: 1593009010
- News link: https://www.theregister.co.uk/2020/06/24/ai_image_tool/
- Source link:
The tool known as [1]PULSE was introduced by a group of researchers from Duke University, and was [2]presented at the virtual Conference on Computer Vision and Pattern Recognition last week. Given a pixelated portrait as input, PULSE searches through computer generated images and picks the best one that it believes is the closest match to the original photo.
All the imaginary images are outputs generated by [3]StyleGAN , a generative adversarial network developed by Nvidia back in 2018. The whole system, essentially, turns a small fuzzy photo into a larger, higher resolution image, a method known as upscaling.
PULSE cannot, however, be used to reveal the true identity of the person hidden in the blurry photo, since it only considers images that have been made up by a generative adversarial network. The people depicted in its results do not exist in real life.
Instead, the tool looks for features like hair length and skin colour in the blurry image and selects a new face dreamed up by StyleGAN that might resemble the person in the original image. Unfortunately, the tool struggles when the obscured photos feature people of colour since it often chooses fake images of people that are white.
Here's a good demonstration of the overall model: The original face on the left may be difficult to make out, but most people would be able to tell that it's former US president Barack Obama.
Given that image, PULSE has selected an image on the right using StyleGAN. The computer-generated face obviously doesn't look anything like Obama at all. In the [4]original image Obama has dark skin, black hair, and brown eyes, but the result is, instead, someone that has white skin, blue eyes, and brown hair.
🤔🤔🤔 [5]pic.twitter.com/LG2cimkCFm — Chicken3gg (@Chicken3gg) [6]June 20, 2020
Robert Ness, a machine learning scientist who teaches workshops at his online platform Altdeep, also found other examples of racial biases when he was toying around with the software. When he fed his own photo and one of Alexandria Ocasio-Cortez, the US Representative for New York's 14th congressional district, the results were both skewed towards Caucasian-looking faces.
An image of [7]@BarackObama getting upsampled into a white guy is floating around because it illustrates racial bias in [8]#MachineLearning . Just in case you think it isn't real, it is, I got the code working locally. Here is me, and here is [9]@AOC . [10]pic.twitter.com/kvL3pwwWe1 — Robert Osazuwa Ness (@osazuwa) [11]June 20, 2020
First, the model blurred both of his original images. Next, PULSE picked images generated by StyleGAN based on the pixelated inputs. "This model demonstrates the same bias issues we've seen in other more commonly used data-driven algorithms like search," he told The Register . "This instance of deep generative modeling just happens to make the issue glaringly obvious."
Do not test a biased model on a biased dataset!
The researchers have acknowledged the issue and believe it stems from existing biases within the StyleGAN model itself. They updated their research [12]paper with a section addressing the issue of racial biases in their work.
The problem is that StyleGAN - trained on [13]70,000 images scraped from Flickr - tends to generate images of white people. In fact, a [14]recent research paper [PDF] examining the demographics of StyleGAN images discovered that it spat out images of white people 72.6 per cent of the time, compared to just 10.1 per cent for black people, and 3.4 per cent for Indian people. PULSE, therefore, is also more likely to choose images of white people since those are the images being generated by StyleGAN.
The team contacted the original researchers from Nvidia to notify them about their issue, but didn't get a response. "NVIDIA takes diversity and inclusion seriously," a spokesperson from the company told The Register .
"We're always striving to create better datasets and algorithms to overcome any existing bias in current models. We are also doing more research on the algorithmic bias in deep-learning models and methods to mitigate them."
Sachit Menon and Alex Damian, both co-authors and recent graduates at the University of Duke, didn't realise this when they decided to carry out the research. They also didn't realise the racial biases in StyleGAN since they tested their tool on another dataset that collects images of celebrities known as [15]CelebA .
"It turns out that 90 per cent of the photos in CelebA are white people," Menon told El Reg . By testing a biased model on a dataset that also contains the same biases, the issue was overlooked.
"We tried [the tool] on me, and I'm Indian," he added. "Sometimes it worked, but a lot of the times it would make me white. And then I found this other paper that showed StyleGAN only makes Indian people 3 per cent of the time. And when we saw this break down, that's when we realised it was biased."
The team used Nvidia's off-the-shelf model and did not train StyleGAN themselves. "In an ideal world, there would be a more balanced dataset, and we could have used a pretrained model that reflected this," said Damian. "If we used a different evaluation dataset that contained more images of people of colour then we would have spotted the issue. That's a big takeaway and a big lesson for us."
The pair said they carried out the project when they were undergraduates and simply selected StyleGAN for their tool because it provides state-of-art results and that CelebA was commonly used to benchmark super-resolution imaging tasks.
"We don't want to point fingers at anyone; implicit biases are systemic issues," they said. "It's important to be aware of these problems that we weren't initially thinking about. We still have a long way to go as a field of being aware of these things."
The research has kickstarted a heated discussion of whether the issue can simply be fixed by using a model that has been trained on more diverse dataset. Facebook's chief AI scientist Yann LeCun believes so, but others like Timnit Gebru, an expert in algorithmic bias, don't think it's as easy as that. ®
I’m sick of this framing. Tired of it. Many people have tried to explain, many scholars. Listen to us. You can’t just reduce harms caused by ML to dataset bias. [16]https://t.co/HU0xgzg5Rt — Timnit Gebru (@timnitGebru) [17]June 21, 2020
Sponsored: [18]Google Security Whitepaper
[1] http://pulse.cs.duke.edu/
[2] http://openaccess.thecvf.com/content_CVPR_2020/html/Menon_PULSE_Self-Supervised_Photo_Upsampling_via_Latent_Space_Exploration_of_Generative_CVPR_2020_paper.html
[3] https://www.theregister.com/2018/12/14/ai_created_photos/
[4] https://www.biography.com/us-president/barack-obama
[5] https://t.co/LG2cimkCFm
[6] https://twitter.com/Chicken3gg/status/1274314622447820801?ref_src=twsrc%5Etfw
[7] https://twitter.com/BarackObama?ref_src=twsrc%5Etfw
[8] https://twitter.com/hashtag/MachineLearning?src=hash&ref_src=twsrc%5Etfw
[9] https://twitter.com/AOC?ref_src=twsrc%5Etfw
[10] https://t.co/kvL3pwwWe1
[11] https://twitter.com/osazuwa/status/1274444300894572546?ref_src=twsrc%5Etfw
[12] http://pulse.cs.duke.edu/
[13] https://github.com/NVlabs/ffhq-dataset
[14] https://dl.acm.org/doi/pdf/10.1145/3334480.3382791
[15] http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html
[16] https://t.co/HU0xgzg5Rt
[17] https://twitter.com/timnitGebru/status/1274809417653866496?ref_src=twsrc%5Etfw
[18] https://go.theregister.com/tl/1956/-8471/google-security-whitepaper?td=wptl1956
All the keys on my keyboard are black.
Your point?
Black keys only
Your point?
A tune in a limited key, of course.
Just as limited as neural networks and often sounds just wrong, like the neural networks.
Mine are a kind of beige/yellow. Think I should probably clean it.
Ha
Didnt Obama use a white man picture of himself to get elected and make the election about electing a black man?
Maybe if these offended are really upset they will go and make one that doesnt have such a problem? Maybe? **tumble weed**
this is not bias
Results are skewed. That is not indicative of bias.
People crying wolf about racial bias will lead to real racism being overlooked. Anyway, at the scale included in the article the image of Obama and not Obama are very comparable indeed.
Re: this is not bias
Bias : "systematic error introduced into sampling or testing by selecting or encouraging one outcome or answer over others" https://www.merriam-webster.com/dictionary/bias
I'd say that definition pretty much covers this case.
Re: this is not bias
By your definition, to be bias the system would have to be rejecting non-white faces in favour of the white ones.
Is that's what's happening?
I've opened both images in an image editor, used a colour picker to select from the cheek, forehead, chin of both images.
- Forehead in the light, both are the same colour.
- Forehead in the shade, both are the same colour.
- Cheek on the left, the selected image is a darker brown than Obama (but similar).
- Chin, centre, Obama is a darker brown than the selected image (but similar)
- Hair, both are the same colour
Where's the bias? Hell, where's the skew?
It doesn't look like it's in the software if people think one image is black and the other white.
Re: this is not bias
Yeah I don't see it either. Both images look equally black.
Re: this is not bias
It's not my definition, its Merriam-Webster's.
The bias is in the training. The GAN learns from the training data what a face looks like. If it sees 90% white faces it will favour white skin in its definition of "face". This isn't a new revelation - it's a well-known phenomenon as the journal article referred to in the Reg article states. It's basically a manifestation of the old maxim "garbage in, garbage out"
Re: this is not bias
It's not bias per se that is the problem but that the Artificial Idiocy consistently fails on relatively simple tasks. Tasks that a child could easily complete with a much higher accuracy rate.
AS - Artificial Stupidity
What we have right now is not Artificial Intelligence, it is Artificial Stupidity.
When you see on a computer science site kids asking if some neural network can solve NP-complete problems, then you realise that the problem is magnified by NS (Natural Stupidity).
An end to police racial bias?
So if the police put in a blurry picture of a BAME suspect then they get a high definition picture of an imaginary white suspect back? Surely this will work to counter-balance any potential discrimination or structural racism and make the world a better place. Or its just a complete piece of junk that should have been tested properly before being released on the world.
What if the image is only very slightly blurry?
Maybe it has some legs as a photo sharpening tool rather than trying to patch up a shoddy surveillance cam image?
Not racially biased, color-blind
Honestly, AIs are not racially biased, they can be "color-blind" (ESPECIALLY when photos are taken in varying lighting conditions). They focus on feature recognition, since they decide on their own what features to look for they can ENTIRELY miss the point sometimes. So, you look at these photos and obviously it's not the same person. You look at FEATURES, and they are surprisingly similar. The eyes in the photos are not brown and blue, to me they both appear black due to lack of resolution. The ears are very similar, the pose is identical, they have the same hair line (including this triangular bit hanging over the forehead), and the lighting they both have a shadow in the top-right corner, the right side of the forehead.
Don't get me wrong, it definitely shows a big problem with facial recognition systems; I'm not a fan of them for privacy reasons either.
One tale of woe regarding AIs.. 10 or 15 years back, the military (don't know if it was US or UK?) was going to test a neural network-based "friend or foe" system. They brought out various airplanes onto the tarmac, took photos to feed in. They train this thing, test it in the wild and it DOES NOT WORK AT ALL. It turns out, most of the friendlys were photographed in the morning, and the rest in the evening, so ALL the AI was basing it's "friend or foe" on was if the plan was lit up from the left side or the right side, it was not looking at what kind of plane it was, the plane markings, etc. at all.
This is not "AI-based image enhancement".
Actual intelligence would recognize the image and fill in the details from memory.
However, an actual intelligence seeing a face which it couldn't recognize, never having seen it before, would do no better than this.
As much as we'd love to extract from an image details which just aren't there, and aren't anywhere else, it can't be done.
Another promise, one which was never believable to begin with, broken.
Yawn.
I always knew computers were racist.
Unlike a piano, all the keys on my keyboard are white.