Seonglae ChoA technique for computing the attribution (contribution) of each input feature in a neural network. The basic idea is to integrate the gradients along a path from a baseline input x′ to the actual input x, measuring how much each input affects the output.
Integrated Gap Gradients
A method for computing the attribution of neurons that contribute to the prediction difference (logit gap) between two groups, using integrated gradients to measure neuron importance
Integrated Gap Gradients