Of course radiologists are freaking out because they don't understand how that's possible (probably because they've never tried) and want to neuter the AI so it's not "biased".
If it's correctly identifying them when they are just blank grey squares, maybe they just overtuned the modeling? I.e. the AI isn't actually detecting anything at all, it just has the dataset memorized.
Granted, I haven't read through all of it, but it seems to me that the grey square filtered are all images that have been fed to the AI before the filters were applied
In other news, AI can detect race from chest x-rays with 99% accuracy and is so good at doing so it "can detect race from images filtered so heavily they are just blank grey squares"
Of course radiologists are freaking out because they don't understand how that's possible (probably because they've never tried) and want to neuter the AI so it's not "biased".
If it's correctly identifying them when they are just blank grey squares, maybe they just overtuned the modeling? I.e. the AI isn't actually detecting anything at all, it just has the dataset memorized.
Granted, I haven't read through all of it, but it seems to me that the grey square filtered are all images that have been fed to the AI before the filters were applied