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

We propose a technique for making CNN-based models more transparent by visualizing the input image regions that are important for predictions from these models- producing visual explanations. Our approach, called Gradient-weighted Class Activation Mapping (Grad-CAM), uses the class-specific gradient information flowing into the final convolutional layer of a CNN to produce a coarse localization map of the regions in the image important for each class. Grad-CAM is a strict generalization of Class Activation Mapping (CAM). Unlike CAM, Grad-CAM is broadly applicable to any CNN-based architectures and needs no re-training. We show how Grad-CAM may be combined with pixel-space visualizations (such as Guided Backprop) to create a high-resolution class-discriminative visualization (Guided Grad-CAM). We generate Grad-CAM and Guided Grad-CAM visualizations to better understand off-the-shelf image classification, image captioning, and visual question answering (VQA) models, including Res-Net based architectures. In the context of image classification models, our visualizations (a) lend insight into model's failure modes, and (b) outperform pixel-space gradient visualizations on the ILSVRC-15 weakly-supervised localization. For image captioning and VQA, our visualizations expose the somewhat surprising insight that common CNN+LSTM models are good at localizing discriminative input image regions despite not being trained on grounded image-text pairs. Finally, through human studies we show that our explanations help users establish trust in the predictions made by deep networks. Interestingly, we find that Guided Grad-CAM helps untrained users successfully discern a stronger deep network from a weaker one even when both make identical decisions. Our code is available at github.com/ramprs/grad-cam/ and a demo is available at gradcam.cloudcv.org. Video of the demo can be found at youtu.be/COjUB9Izk6E.