Use HiResCAM instead of Grad-CAM for faithful explanations of convolutional neural networks

Explanation methods facilitate the development of models that learn concepts and avoid exploiting spurious correlations. We illustrate a unrecognized limitation of the popular neural network explanation Grad-CAM: as a side effect of the gradient averaging step, Grad-CAM highlights locations the model did not actually use. To solve this, we propose HiResCAM, a novel class-specific explanation method that is to highlight only the locations the model used to make each. We prove that HiResCAM is a generalization of CAM and explore the between HiResCAM and other gradient-based explanation methods. on PASCAL VOC 2012, including crowd-sourced evaluations, illustrate while HiResCAM's explanations faithfully reflect the model, Grad-CAM often the attention to create bigger and smoother visualizations. Overall, work advances convolutional neural network explanation approaches and may in the development of trustworthy models for sensitive applications.