Grad-CAM++: Generalized Gradient-Based Visual Explanations for Deep Convolutional Networks

Abstract

Over the last decade, Convolutional Neural Network (CNN) models have been\nhighly successful in solving complex vision problems. However, these deep\nmodels are perceived as "black box" methods considering the lack of\nunderstanding of their internal functioning. There has been a significant\nrecent interest in developing explainable deep learning models, and this paper\nis an effort in this direction. Building on a recently proposed method called\nGrad-CAM, we propose a generalized method called Grad-CAM++ that can provide\nbetter visual explanations of CNN model predictions, in terms of better object\nlocalization as well as explaining occurrences of multiple object instances in\na single image, when compared to state-of-the-art. We provide a mathematical\nderivation for the proposed method, which uses a weighted combination of the\npositive partial derivatives of the last convolutional layer feature maps with\nrespect to a specific class score as weights to generate a visual explanation\nfor the corresponding class label. Our extensive experiments and evaluations,\nboth subjective and objective, on standard datasets showed that Grad-CAM++\nprovides promising human-interpretable visual explanations for a given CNN\narchitecture across multiple tasks including classification, image caption\ngeneration and 3D action recognition; as well as in new settings such as\nknowledge distillation.\n

Keywords

Affiliated Institutions

• Cisco Systems (Norway)
• Indian Institute of Technology Hyderabad IN

Related Publications