Grad-CAM is a model-dependent, gradient-based saliency method for convolutional neural networks (CNNs).
Grad-CAM identifies continuous input regions that are important to the model's output towards the target class. It computes feature importance by extracting the feature maps from an intermediate convolutional layer (typically the last convolutional layer) and weighting them by the gradient of the target output with respect to that layer. The weighted feature maps are summed to obtain a single map, passed through a ReLU function to remove negatively contributing values, and upsampled to the original input dimensions.
Developed by: Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra at Georgia Institute of Technology.
References:
Implementations and Tutorials:
- Original GitHub Repository: ramprs/grad-cam
- PyTorch Integration via Captum: Captum Grad-CAM
- Keras Integration: Keras Grad-CAM Tutorial
Example: The Grad-CAM saliency map (right) on an ImageNet image for the class boxer (left) using a VGG-16. This example is from Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization.
Grad-CAM is deterministic.
Grad-CAM relies on two hyperparameters: the interpolation method and the convolutional layer.
- The
interpolation methodupsamples the feature map into the input feature dimensions. - The
convolutional layerdetermines which feature maps to use. Typically, the last convolutional layer is used, but any convolutional layer can be used.
Grad-CAM requires a differentiable model with convolutional layers and access to the gradients.
Computing Grad-CAM takes on the order of
Grad-CAM outputs the positive attributions of the gradient-weighted feature maps from an internal convolutional layer. Interpreting it requires an understanding of convolutional models and model gradients.
🟨 Deletion: Grad-CAM's deletion performance is inconclusive. When evaluated on ResNet50 and VGG16 ImageNet models, Grad-CAM performs better than sliding window saliency but worse than RISE and LIME. In subsequent evaluations on ResNet50 and CUB-200-2011, Grad-CAM performs similarly to RISE and worse than Ablation CAM, Grad-CAM++, and Score-CAM.
🟨 Insertion: Grad-CAM's insertion performance is inconclusive. When evaluated on ResNet50 and VGG16 ImageNet models, Grad-CAM performs better than sliding window saliency but worse than RISE. It performs worse than LIME using a ResNet50 model and on par with LIME using a VGG16. In subsequent evaluations on ResNet50 and CUB-200-2011, Grad-CAM performs worse than RISE, Ablation CAM, Grad-CAM++, and Score-CAM.
🟢 Data Randomization: Grad-CAM saliency changes appropriately when the model is trained on perturbed data labels. Evaluated on MNIST and Fashion MNIST using CNN and MLP models.
🟢 Model Contrast Score: Grad-CAM acheives the highest model contrast score compared to vanilla gradients, SmoothGrad, integrated gradients, integrated gradients with SmoothGrad, guided backpropagation, and guided Grad-CAM. Evaluated on the BAM image dataset.
🟢 Cascading Model Parameter Randomization: Grad-CAM saliency changes appropriately as the model is progressively randomized. Evaluated on an ImageNet Inception V3.
🟢 Independent Model Parameter Randomization: Grad-CAM saliency changes appropriately as the model layers are independently randomized. Evaluated on an ImageNet Inception V3.
🟢 Model Weight Randomization: Grad-CAM saliency differs appropriately between a fully trained and fully randomized model. Evaluated on SIIM-ACR Pneumothorax and RSNA Pneumonia medical images.
🟨 Repeatability: Grad-CAM's repeatability is similar to/slightly better than a random baseline. Evaluated on SIIM-ACR Pneumothorax and RSNA Pneumonia medical images.
🟨 Reproducibility: Grad-CAM's reproducibility is inconclusive. Its saliency is somewhat consistent between two models with different architectures trained on the same data but performs worse than a segmentation model. Evaluated on SIIM-ACR Pneumothorax and RSNA Pneumonia medical images.
🟨 Sparsity: Grad-CAM's sparsity ratio is 5.28. It has lower sparsity than Ablation-CAM, Grad-CAM++, RISE, and Score-CAM. Evaluated on a ResNet50 model and CUB-200-2011 dataset.
🟥 Localization Utility: Grad-CAM fails localization utility. Its saliency values overlap less with the ground truth than a random model. Evaluated on SIIM-ACR Pneumothorax and RSNA Pneumonia medical images.
🟥 Luminosity Calibration: Grad-CAM saliency values reflect the impact on the target score as much as random saliency. Evaluated on a ResNet50 model and CUB-200-2011 dataset.
🟨 Mean IoU: Grad-CAM saliency has a higher mean IoU than other saliency methods (integrated gradients, Grad-CAM++, Eigen-CAM, DeepLift, LRP, and Occlusion) but a lower mean IoU than human localization. Evaluated using CNNs on CheXpert chest x-ray images.
🟨 The Pointing Game: Grad-CAM's most salient feature was in the ground truth region as many times as other saliency methods, but less than human localization. Evaluated using CNNs on CheXpert chest x-ray images by Benchmarking saliency methods for chest X-ray interpretation.
@inproceedings{grad-cam,
author = {Ramprasaath R. Selvaraju and Michael Cogswell and Abhishek Das and Ramakrishna Vedantam and Devi Parikh and Dhruv Batra},
title = {Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization},
booktitle = {International Conference on Computer Vision ({ICCV})},
publisher = {{IEEE} Computer Society},
year = {2017},
}