How much is learned from an image?

Hey guys I was wondering if it was possible to get a score of how much was learned from an image. This could help discard useless images or noisy ones where nothing is learned.


Any ideas? Couldn’t find anything online

The methods they use in the following paper might help perhaps.

but what if the images you mark as “not worthy” might by limitations of your model not the dataset ?

I don’t mind! it’ll be model dependent. The question is, for my model, what are the important images?