3 ms·

The "white Obama" pictures didn't even seem biased. They took a brightly lit photo of Obama and the output face has the same skin tone. Compare Obama's skin to

The "white Obama" pictures didn't even seem biased. They took a brightly lit photo of Obama and the output face has the same skin tone. Compare Obama's skin tone in this photo: https://video-images.vice.com/test-uploads/articles/5ef216e9aea1330097d9ec64/lede/1592924226088-obama.jpeg?crop=0.9075439591605219xw:1xh;center,center&resize=640 https://video-images.vice.com/test-uploads/articles/5ef216e9...:* With other photos. E.g: https://www.obama.org/wp-content/uploads/ar_obama_sq.jpg https://www.obama.org/wp-content/uploads/ar_obama_sq.jpg

godelski 5y ago
Bias is an overloaded word. To be clear, there is bias (there's always bias). Every model always has bias, we just want to understand that to understand the limitations of it. But yes, I agree that the reconstruction is decently faithful. I'm sure an L2 loss isn't that high when comparing the images. Part of what I was trying to say is that there is a clear human bias when evaluating the reconstruction. The problem here is our prior. We humans already know that the compressed image is Obama because he is a well known face. So we're judging based on that prior. So there is some irony when talking about model bias when you don't recognize the prior bias in evaluation. We should analyze the bias of the model (I mean there's problems with the reconstruction other than the race issue, like the teeth), but we have to also be careful when analyzing because we also have a bias (which we may or may not want to impart upon the model). The thing that gets me about these bias people in ML is that it is easy to recognize and identify bias. It is much harder to solve them. There seems to be two camps. Those that just want to identify and those that work on resolving them. The former typically say they work in bias and the latter typically says they work on long tail data (or sometimes few/low shot learning, or even metalearning). Yeah, we should recognize biases and discuss them, but just yelling at people doesn't help develop the necessary mathematical frameworks to resolve the issues.
sangnoir 5y ago
You can't can't use a use a dropper and compare color hex/HSL values on different pictures with different lighting and say "Yup - that's the same", otherwise SFX wouldn't need color grading[1]. The second picture seems to be from a very brightly-lit environment - to the extent of washing out his blue shirt (collar), but the first picture gives no such cues, and appears less well-lit (virtually). AI models have no implicit knowledge of environmental lighting, and if the training data is exceedingly made of brightly lit images, may end up generating images that belong in well-lit environments (fairer-looking skin tones), but in low-light contexts. When looking at these generated images, the human eye does what it does best: determine relevant gamma-correction using all available cues. [1]. Assets rendered in the same color under all scene lighting conditions.

godelski 5y ago

Bias is an overloaded word. To be clear, there is bias (there's always bias). Every model always has bias, we just want to understand that to understand the limitations of it. But yes, I agree that the reconstruction is decently faithful. I'm sure an L2 loss isn't that high when comparing the images. Part of what I was trying to say is that there is a clear human bias when evaluating the reconstruction. The problem here is our prior. We humans already know that the compressed image is Obama because he is a well known face. So we're judging based on that prior. So there is some irony when talking about model bias when you don't recognize the prior bias in evaluation. We should analyze the bias of the model (I mean there's problems with the reconstruction other than the race issue, like the teeth), but we have to also be careful when analyzing because we also have a bias (which we may or may not want to impart upon the model). The thing that gets me about these bias people in ML is that it is easy to recognize and identify bias. It is much harder to solve them. There seems to be two camps. Those that just want to identify and those that work on resolving them. The former typically say they work in bias and the latter typically says they work on long tail data (or sometimes few/low shot learning, or even metalearning). Yeah, we should recognize biases and discuss them, but just yelling at people doesn't help develop the necessary mathematical frameworks to resolve the issues.

sangnoir 5y ago

You can't can't use a use a dropper and compare color hex/HSL values on different pictures with different lighting and say "Yup - that's the same", otherwise SFX wouldn't need color grading[1]. The second picture seems to be from a very brightly-lit environment - to the extent of washing out his blue shirt (collar), but the first picture gives no such cues, and appears less well-lit (virtually). AI models have no implicit knowledge of environmental lighting, and if the training data is exceedingly made of brightly lit images, may end up generating images that belong in well-lit environments (fairer-looking skin tones), but in low-light contexts. When looking at these generated images, the human eye does what it does best: determine relevant gamma-correction using all available cues. [1]. Assets rendered in the same color under all scene lighting conditions.