Grad CAM

Grad-CAM is a technique used to visualize and interpret the decision-making process of convolutional neural networks (CNNs), primarily focusing on identifying which image regions are most influential in a model's prediction. Current research explores Grad-CAM's application across diverse CNN architectures, including ResNets and more recent models, and investigates improvements to its robustness and faithfulness, addressing issues like gradient saturation and sensitivity to baseline parameters. This work is significant for enhancing the transparency and trustworthiness of AI models, particularly in high-stakes applications like medical image analysis and other domains requiring explainable AI.

Papers

November 22, 2022

Explaining YOLO: Leveraging Grad-CAM to Explain Object Detections
Armin Kirchknopf, Djordje Slijepcevic, Ilkay Wunderlich, Michael Breiter, Johannes Traxler, Matthias Zeppelzauer
Data Detection Object Detection High Explainability YOLO Object YOLO Shake Hand Grad CAM Attribution Based

May 22, 2022

April 13, 2022

Baseline Computation for Attribution Methods Based on Interpolated Inputs
Miguel Lerma, Mirtha Lucas
Neural Network Interpolation Regime Attribution Method Grad CAM Class Activation Mapping Baseline Algorithm

April 11, 2022

Generalizing Adversarial Explanations with Grad-CAM
Tanmay Chakraborty, Utkarsh Trehan, Khawla Mallat, Jean-Luc Dugelay
Black Box Grad CAM General Adversarial Class Activation Mapping

Grad CAM

Papers

Explaining YOLO: Leveraging Grad-CAM to Explain Object Detections

Visual Explanations from Deep Networks via Riemann-Stieltjes Integrated Gradient-based Localization

Grad-CAM++ is Equivalent to Grad-CAM With Positive Gradients

Baseline Computation for Attribution Methods Based on Interpolated Inputs

Generalizing Adversarial Explanations with Grad-CAM