3 ms·
I have a lot of sympathy for the researchers in this case. The picture given in the article shows a skin tone that is certainly within the range of skin tones
by liability 7y ago
I have a lot of sympathy for the researchers in this case. The picture given in the article shows a skin tone that is certainly within the range of skin tones that you could reasonably expect from caucasians (I am darker than her, on account of being 5/8ths Sicilian), so differentiating her 'race' is a much more subtle matter than checking the tone of her skin. I expect to handle this case gracefully, the system needs to be specifically engineered to detect race, but that is a live wire for reasons that should be obvious.
Could they have done better? I'm sure. But I don't think bullying the researchers is a productive step towards better performance.
- visarga 7y agoIt was trained on renaissance paintings. How many of those feature Asians or black people? It wasn't a portrait generator, it was a 'renaissance conditioned portrait generator'. If it were trained on Ukiyo E, it would have made everyone look Asian. GANs have nothing against specific races.
- liability 7y agoYes, that was clear from the article. What I'm saying is that if they wanted to make this thing preserve race, I expect they'd need to give it an explicit notion of race. If the suggested problem is that they trained it to one art style rather than every art style, I reject that. There is nothing wrong with a narrow artistic focus. Furthermore, I expect that if the researchers trained one model for each 'race' using an appropriate art style for each, they would still probably get complaints about even having different models, or for the models 'failing' when used on the 'wrong' race, or when somebody with ambiguous 'racial characteristics' tried to use it.