Towards robust visual explanations for deep convolutional networks with weight-wise perturbations

Explaining the complicated mechanisms underlying the image classification capabilities of deep Convolutional Neural Network (CNN) models, continue to pose huge challenges within the field of computer vision. To address this concern, numerous interpretability methods have been devised, aimed at clarifying the image classification process. These methods include approaches such as sensitivity maps, which involve computing the gradients of class outputs with respect to input images, and techniques like class activation mapping (CAM). Furthermore, the incorporation of noise into input images has emerged as an effective strategy for augmenting visualization quality and removing noise. In this paper, we propose two key contributions: the introduction of a novel approach that injects noise into network weights to enhance visualization, involving image gradient updates and average gradient computations; and a new indicator for evaluating interpretability - the center of gravity, and comprehensive experiments were conducted on multiple datasets and different deep neural network models. In the subsequent experimental sections, we demonstrate that our method achieves superior visualization quality and can be combined with other interpretability methods to enhance their performance.