That's true, I know that you can view what "excites" specific neurons. I still think the is a problem: this only shows whenenever you give an image what it uses

That's true, I know that you can view what "excites" specific neurons. I still think the is a problem: this only shows whenenever you give an image what it uses from that image to classify the object, but when a new, unseen image is given it's really hard to tell beforehand what the network will actually look like and which pixels would be considered. Yes, you could look at the features and see which patterns appear there, but for large networks those are thousands or millions, so it's still hard to tell what the result will be, unless you actually use the new image as an input for the network.