The moment you need ML magic to magnify you are basically telling the ML "guess what is there based on what we can see and the training database you where trained on". Also due to the nature of such algorithm they do not have any form of common sense or "realistic general purpose" world view, and tend to be black box so it's hard to check if they learned something "stupid"/"ridiculous".
So it's totally possible that somehow it learned to place knifes into the hyper-magnified hands if they have a specific skin color and a turquoise sleeve on a picture with a color scheme as if it's a cloudy day. Or similar arbitrary things. Sure it's supper unlikely but not impossible and there is not good way to find such "bugs". Worse multiple ML systems trained on the same data might share some mistakes and there are just that many huge comprehensive datasets.
OK sure but the distribution of natural images is highly redundant ie pixels are not independent and there is structure to exploit. It's not magic obviously but done properly it's a pretty reasonable thing
> OK sure but the distribution of natural images is highly redundant ie pixels are not independent and there is structure to exploit.
Sure, but that depends very strongly on the number of natural images you trained on and the specificity of the situation you present to the algorithm. Maybe one such structure in the training set just happens to be people holding knives given "a specific skin color and a turquoise sleeve on a picture with a color scheme as if it's a cloudy day".
Sure, if it is an option applied after the fact by the user editing the photo. Any camera that does this automatically is utterly broken. Otherwise where is the line? Some sharpening? Adding bigger boobs by default? Make people less black?
The moment you need ML magic to magnify you are basically telling the ML "guess what is there based on what we can see and the training database you where trained on". Also due to the nature of such algorithm they do not have any form of common sense or "realistic general purpose" world view, and tend to be black box so it's hard to check if they learned something "stupid"/"ridiculous".
So it's totally possible that somehow it learned to place knifes into the hyper-magnified hands if they have a specific skin color and a turquoise sleeve on a picture with a color scheme as if it's a cloudy day. Or similar arbitrary things. Sure it's supper unlikely but not impossible and there is not good way to find such "bugs". Worse multiple ML systems trained on the same data might share some mistakes and there are just that many huge comprehensive datasets.