We propose a technique for producing "visual explanations" for decisions from a large class of CNN-based models, making them more transparent. Our approach — Gradient-weighted Class Activation Mapping (Grad-CAM) — uses the gradients of any target concept flowing into the final convolutional layer to produce a coarse localization map highlighting the important regions in the image for predicting the concept. Grad-CAM is applicable to a wide variety of CNN model-families including CNNs with fully-connected layers (e.g., VGG), CNNs used for structured outputs (e.g., captioning), CNNs used in tasks with multi-modal inputs (e.g., VQA), without any architectural changes or re-training. We combine Grad-CAM with existing fine-grained visualizations to create a high-resolution class-discriminative visualization, Guided Grad-CAM. We show that these visualizations help identify dataset bias, build trust in model predictions, and enable model diagnosis.

Deep neural networks have enabled unprecedented breakthroughs across tasks in computer vision, but their lack of decomposability into intuitive and understandable components makes them hard to interpret. As these models are increasingly deployed in real-world applications, the need for transparency and interpretability becomes paramount. A "good" visual explanation should be class-discriminative (localizing the category in the image) and high-resolution (capturing fine-grained detail). Prior visualization approaches achieve one but not both: pixel-space gradient methods (Guided Backpropagation, Deconvolution) are high-resolution but not class-discriminative; CAM is class-discriminative but requires specific architectures with global average pooling layers directly before the softmax.

Class Activation Mapping (CAM) identifies discriminative image regions by using weights from the global average pooling layer to linearly combine feature maps. While effective, CAM is restricted to architectures ending with convolutional feature maps followed by global average pooling and a softmax layer. Gradient-based methods like Guided Backpropagation and Deconvolution provide pixel-level detail but are not class-discriminative — they highlight the same regions regardless of the target class. We prove that Grad-CAM is a strict generalization of CAM: when applied to the specific architecture CAM requires, Grad-CAM produces identical results, but Grad-CAM can be applied to any CNN architecture without modification.

Grad-CAM computes the gradient of the score for class c (y^c) with respect to feature map activations A^k of the last convolutional layer. These gradients are global-average-pooled to obtain the neuron importance weights alpha_k^c. The Grad-CAM localization map is then: L_Grad-CAM^c = ReLU(sum_k(alpha_k^c * A^k)). The ReLU is applied because we are only interested in features that have a positive influence on the class of interest — pixels whose intensity should be increased to increase y^c. The resulting map has the same spatial dimensions as the last convolutional feature maps (e.g., 14x14 for VGG-16) and is upsampled to image resolution for visualization.

While Grad-CAM provides class-discriminative localization, it is coarse (e.g., 14x14). To achieve both high-resolution and class-discriminative visualizations, we fuse Grad-CAM with Guided Backpropagation via element-wise multiplication: the Grad-CAM map is first upsampled to the input resolution, then multiplied with the Guided Backpropagation map. This produces Guided Grad-CAM, which retains fine-grained pixel detail from Guided Backpropagation while being modulated by the class-discriminative coarse map from Grad-CAM. In human studies, Guided Grad-CAM achieved 61.23% accuracy in class discrimination tests, compared to Guided Backpropagation's 44.44%.

We evaluate Grad-CAM across multiple dimensions. For weakly-supervised localization on ILSVRC-15, Grad-CAM achieves 56.51% top-1 localization error on VGG-16, outperforming CAM (57.20%) while maintaining classification performance. On weakly-supervised segmentation (PASCAL VOC 2012), replacing CAM with Grad-CAM as seeds improved Intersection over Union from 44.6 to 49.6. In the pointing game evaluation, Grad-CAM achieved 70.58% accuracy versus c-MWP's 60.30%. For faithfulness evaluation, Grad-CAM's rank correlation with occlusion sensitivity reached 0.254, outperforming Guided Backpropagation (0.168), CAM (0.208), and c-MWP (0.220), demonstrating better fidelity to actual model behavior despite being more interpretable.

We demonstrate Grad-CAM's utility for model diagnosis and bias detection. In a doctor/nurse classification task, Grad-CAM revealed that the model learned gender stereotypes by focusing on faces and hairstyles rather than medical equipment. The training data was gender-imbalanced: 78% male doctors, 93% female nurses, causing poor generalization (82% test accuracy). After adding balanced gender representation, accuracy improved to 90% with the model correctly focusing on relevant features. Grad-CAM also revealed that "seemingly unreasonable predictions have reasonable explanations" — misclassifications often occurred because the model focused on legitimate but insufficient features. For adversarial robustness, despite networks assigning near-zero probability to true categories after perturbations, Grad-CAM correctly localized the original objects, suggesting adversarial attacks affect the classification layer more than feature representations.

We proposed Grad-CAM, a class-discriminative localization technique that is applicable to any CNN-based architecture without architectural changes or retraining. By combining Grad-CAM with Guided Backpropagation, Guided Grad-CAM achieves both high-resolution and class-discriminative visualizations. We demonstrated that these visualizations are faithful to the model, help identify dataset biases, build appropriate trust, and enable diagnosis across multiple architectures and tasks including image classification, captioning, and visual question answering. Grad-CAM represents a step toward making deep learning models more transparent and trustworthy.