I am currently learning convolutional neural networks, and its visualization algorithm Grad-CAM. I implemented them in keras, and the results looks decent.
My question is that when calculating guided Grad-CAM (multiplication of Grad-CAM value and guided back-propagation value), the result have both positive and negative score for the image.
My initial thought was that positive guided Grad-CAM value indicates that the presence of the respective pixel results in increased score in the respective class, and negative value indicates the absence of the pixel results in increased score.
Is this interpretation correct?
Sincerely,