Grad-CAM is a technique for generating visual explanations for decisions made by deep neural networks, making them more transparent and explainable. It uses gradients flowing into the final convolutional layer to produce a coarse localization map highlighting important image regions for predicting a concept. Unlike previous methods, Grad-CAM is applicable to a wide variety of CNN model families without architectural changes or re-training. It combines with existing fine-grained visualizations to create Guided Grad-CAM, which is both high-resolution and class-discriminative. Grad-CAM has been applied to image classification, image captioning, and visual question answering (VQA) models, including ResNet-based architectures. In image classification, Grad-CAM visualizations help identify failure modes, outperform previous methods on weakly-supervised localization tasks, are robust to adversarial perturbations, and help achieve model generalization by identifying dataset bias. For image captioning and VQA, Grad-CAM shows that non-attention based models can learn to localize discriminative image regions. Grad-CAM can identify important neurons and combine them with neuron names to provide textual explanations for model decisions. Human studies show that Grad-CAM helps users establish appropriate trust in predictions from deep networks and discern a 'stronger' network from a 'weaker' one even when both make identical predictions. Grad-CAM is a generalization of CAM and is applicable to a broader range of CNN model families. It is class-discriminative and can be used to generate visual explanations for any CNN-based network without altering its architecture. Grad-CAM has been evaluated for localization and faithfulness to the model, where it outperforms baselines. It has been applied to existing top-performing classification, captioning, and VQA models, showing that current CNNs can have reasonable explanations for seemingly unreasonable predictions. Grad-CAM visualizations also help diagnose failure modes by uncovering biases in datasets, which is important for fair and bias-free outcomes. Grad-CAM is also used to generate textual explanations for model decisions by combining neuron importance with neuron names. It has been applied to vision and language models, including image captioning and VQA. Grad-CAM helps interpret image captioning models by finding spatial support regions for captions in images. It has been shown to outperform other visualization techniques in terms of localization accuracy and class-discriminativeness. Grad-CAM is also used to generate high-resolution visual explanations for image captioning and VQA tasks, which are more accurate and detailed than previous methods. Grad-CAM has been shown to be robust to adversarial noise and can help detect and remove biases in training datasets.Grad-CAM is a technique for generating visual explanations for decisions made by deep neural networks, making them more transparent and explainable. It uses gradients flowing into the final convolutional layer to produce a coarse localization map highlighting important image regions for predicting a concept. Unlike previous methods, Grad-CAM is applicable to a wide variety of CNN model families without architectural changes or re-training. It combines with existing fine-grained visualizations to create Guided Grad-CAM, which is both high-resolution and class-discriminative. Grad-CAM has been applied to image classification, image captioning, and visual question answering (VQA) models, including ResNet-based architectures. In image classification, Grad-CAM visualizations help identify failure modes, outperform previous methods on weakly-supervised localization tasks, are robust to adversarial perturbations, and help achieve model generalization by identifying dataset bias. For image captioning and VQA, Grad-CAM shows that non-attention based models can learn to localize discriminative image regions. Grad-CAM can identify important neurons and combine them with neuron names to provide textual explanations for model decisions. Human studies show that Grad-CAM helps users establish appropriate trust in predictions from deep networks and discern a 'stronger' network from a 'weaker' one even when both make identical predictions. Grad-CAM is a generalization of CAM and is applicable to a broader range of CNN model families. It is class-discriminative and can be used to generate visual explanations for any CNN-based network without altering its architecture. Grad-CAM has been evaluated for localization and faithfulness to the model, where it outperforms baselines. It has been applied to existing top-performing classification, captioning, and VQA models, showing that current CNNs can have reasonable explanations for seemingly unreasonable predictions. Grad-CAM visualizations also help diagnose failure modes by uncovering biases in datasets, which is important for fair and bias-free outcomes. Grad-CAM is also used to generate textual explanations for model decisions by combining neuron importance with neuron names. It has been applied to vision and language models, including image captioning and VQA. Grad-CAM helps interpret image captioning models by finding spatial support regions for captions in images. It has been shown to outperform other visualization techniques in terms of localization accuracy and class-discriminativeness. Grad-CAM is also used to generate high-resolution visual explanations for image captioning and VQA tasks, which are more accurate and detailed than previous methods. Grad-CAM has been shown to be robust to adversarial noise and can help detect and remove biases in training datasets.