Loss function for one manipulating one coordinate of the prediction

[VergiliusShao]

We want to find the set of pixels for changing the value of one pre-selected output coordinate. So if the prediction of an image is (0.2,0.3,0.4,0.5)
. We might want explanations that the prediction could be (200, 0.3,0.4,0.5). We can use a loss function
describing the similarity of the first score and the dissimilarity of the rest scores
and minimise it.
Let's say the prediction of some sample x is f(x)
and x_0 is the original input.
What we can do is minimise sim(f(x_0)_1,f(x_0+x)_1)+dis_sim(f(x_0)_2,f(x_0+x)_2)+,...,