Get the “sensitivity” of each unit with respect to perturbations of the bias

As the paper “Notes on Convolutional Neural Networks” tells, I just guessed that the “sensitivity” is a very important factor in deep learning.
My question is how can we get the “sensitivity”: δ = ∂E / ∂b = ∂E / ∂u in PyTorch?
Is there some efficient function? Maybe it’s something like this one.