Exploring gradient-based prediction explanations
Most up-to-date neural network-based explanation techniques today are variations of using the gradients that can be obtained through backpropagation. Gradient-based explanations for neural network models work because they rely on the fundamental principle of how the weights in a neural network are updated during the training process using backpropagation. During backpropagation, the partial derivatives of the loss function concerning the weights in the network are calculated, which gives us the gradient of the loss function concerning the weights.
This gradient provides us with a measure of how much the input data contributes to the overall loss. Remember that gradients measure the sensitivity of the input value concerning the loss function. This means it provides the degree of fluctuation of the predictions when you modify the specific input value, which represents the importance of the input data. Input data can be chosen to be...
