# Grad-CAM: Why did you say that? Visual Explanations from Deep Networks via Gradient-based Localization

**Authors:** Ramprasaath R. Selvaraju, Abhishek Das, Ramakrishna Vedantam, Michael Cogswell, Devi Parikh, Dhruv Batra  
**arXiv:** [1610.02391](https://arxiv.org/abs/1610.02391)  
**Date:** 2016-10-07  
**Summary page:** https://www.summarizepaper.com/en/paper/1610.02391v1/

## Key points

- Authors propose Gradient-weighted Class Activation Mapping (Grad-CAM) to enhance transparency of CNN-based models
- Grad-CAM generates coarse localization map of significant regions in image using class-specific gradient information
- Unlike Class Activation Mapping (CAM), Grad-CAM does not require re-training and can be applied to any CNN-based architecture
- Guided Grad-CAM combines Grad-CAM with existing pixel-space visualizations for high-resolution class-discriminative explanations
- Effectiveness of Grad-CAM and Guided Grad-CAM demonstrated in image classification, image captioning, and VQA models
- Visualizations provide insights into failure modes of image classification models and outperform other pixel-space gradient visualizations on weakly supervised localization task
- Common CNN + LSTM models used for image captioning and VQA can effectively localize discriminative input image regions despite not being trained on grounded image-text pairs
- Human studies show that untrained users can discern "stronger" deep network from "weaker" one with the help of Guided Grad-CAM
- Overall, Grad-CAM makes CNN-based models more transparent and understandable by visualizing important regions in input images

## Summary

In their paper titled "Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization," authors Ramprasaath R. Selvaraju, Abhishek Das, Ramakrishna Vedantam, Michael Cogswell, Devi Parikh, and Dhruv Batra propose a technique called Gradient-weighted Class Activation Mapping (Grad-CAM) to enhance the transparency of Convolutional Neural Network (CNN)-based models. The goal is to visualize the important regions in an input image that contribute to the model's predictions. Grad-CAM utilizes class-specific gradient information from the final convolutional layer of a CNN to generate a coarse localization map of significant regions in the image. Unlike its predecessor, Class Activation Mapping (CAM), Grad-CAM does not require re-training and can be applied to any CNN-based architecture. The authors also introduce Guided Grad-CAM, which combines Grad-CAM with existing pixel-space visualizations to create high-resolution class-discriminative visual explanations. The authors demonstrate the effectiveness of Grad-CAM and Guided Grad-CAM in understanding image classification, image captioning, and visual question answering (VQA) models. In terms of image classification models, these visualizations provide insights into their failure modes by revealing reasonable explanations for seemingly unreasonable predictions. Additionally, on the ILSVRC-15 weakly supervised localization task, Grad-CAM outperforms pixel-space gradient visualizations such as Guided Backpropagation and Deconvolution. Surprisingly, the authors find that common CNN + LSTM models used for image captioning and VQA can effectively localize discriminative input image regions despite not being trained on grounded image–text pairs. This insight is revealed through their visualizations. To evaluate if Guided Grad–CAM explanations help users establish trust in deep network predictions, human studies are conducted. Interestingly, it is shown that untrained users can successfully discern a "stronger" deep network from a "weaker" one even when both networks make identical predictions with the help of Guided Grad–CAM. Overall, the proposed Grad–CAM technique provides a means to make CNN–based models more transparent and understandable by visualizing important regions in input images. The authors' experiments and human studies demonstrate the effectiveness and potential applications of Grad–CAM in various domains such as image classification, image captioning, and VQA.