Integrated Gradient(IG) computes the gradient of the model’s prediction output to its input features and requires no modification to the original deep neural network. IG can be applied to any differentiable model like image, text, or structured data.
To calculate the Sensitivity, we establish a Baseline image as a starting point. We then build a sequence of images which we interpolate from a baseline image to the actual image to calculate the integrated gradients.
Implementation invariance is satisfied when two functionally equivalent networks have identical attributions for the same input image and the baseline image.
Two networks are functionally equivalent when their outputs are equal for all inputs despite having very different implementations.