A type of Saliency Map that works by calculating the gradient of the output of a CNN with respect to the outer-most Convolution layer.
The importance of each pixel is then found by evenly distributing the gradient to the pixels that correspond to it in the Convolution. This means that each of the pixels that can be backtracked from the element of the convolution each gets of the gradient value. This requires backtracking to see which pixels are part of which element in the convolution.