https://arxiv.org/pdf/1610.02391
(Guided) Grad-CAM is a technique for improving the interpretability of deep learning models for image classification, captioning orquestion answring tasks. Combining Grad-CAM and Guided Backpropagation to visualize which parts of a convolutional neural network contribute to classification. It can visualize which part of the convolutional neural network contributes to the tasks.
A method of measuring how large an impact a pixel had on the judgment value. The degree of influence is determined by using the gradient to the output label with respect to feature map A.
Guided Grad-CAM modifies the feature map by multiplying the weighted feature map obtained by Grad-CAM with the accurate gradient information obtained by Guided Backpropagation. The modified feature map is then mapped onto the input image and visualized to more accurately identify regions of the input image that contribute to classification.
