I am researching attribution methods in computer vision literature to better understand how a CNN model arrives at its predictions. I have come across the terms class activation map and saliency map, and my understanding is that these are approaches that highlight which pixels/regions in the input image have the largest impact on the output prediction. Based off how they are described in articles and lectures I've come across, it seems like these two terms are interchangeable? I wanted to verify if that is indeed correct and if not, what is the difference? Thanks!

  • $\begingroup$ The difference is in the architecture and the way those maps are computed: medium.com/@mrsalehi/… $\endgroup$ Commented Jan 18, 2021 at 13:43

1 Answer 1


While 'traditionally' and formally, most often saliency maps refer to visualizations of the gradient in the network, activation maps focus more on the global importance of particular region to the final outcome.

However, nowadays the two terms are often used interchangeably, so it is good to define the way the method used to generate the map, e.g. Grad-CAM.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.