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

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