“Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization” was submitted to arXiv on October 7, 2016 by Ramprasaath Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. It was presented at ICCV 2017 and later expanded in the International Journal of Computer Vision. Grad-CAM stands for Gradient-weighted Class Activation Mapping, and it produces a heatmap showing where in an image a convolutional network “looked” when making a decision.
The method works on the network’s final convolutional layer, where each channel acts as a detector for some visual pattern. Grad-CAM measures how much each channel matters to a target class by taking the gradient of the class score with respect to that channel’s feature map, averages those gradients into a weight per channel, and combines the channels into a single coarse map. The result, overlaid on the original photo, lights up the regions that drove the prediction - the dog’s face for “dog,” the stripes for “zebra.”
A key advantage over the earlier CAM method is that Grad-CAM needs no changes to the network architecture and no retraining, so it applies to almost any existing CNN, including ones doing captioning or visual question answering. The authors used it to expose failure modes, such as a model that classified images by exploiting dataset biases rather than the intended object, and to show that more accurate models tend to attend to more reasonable regions.
Why business readers should care: in medical imaging, quality inspection, and content moderation, a heatmap showing what the model focused on is often the difference between a system experts will trust and one they will not.