02·GradCAM: gradient-weighted class activation maps

Introduction

GradCAM (Selvaraju et al. 2017) is probably the most widely used method in practice for visualizing "where the network looks" when making a classification decision. It generalizes the earlier Class Activation Map (CAM, Zhou et al. 2016) to any CNN architecture — it does not require a GAP+Linear head. The algorithm is elegant: for a chosen class c we compute the gradient of the logit y^c with respect to the activation maps A^k of a specific (usually last) convolutional layer, average that gradient over spatial dimensions to obtain weights α_k^c, and then compute a weighted sum of activation maps with a ReLU applied: L_GradCAM^c = ReLU(Σ_k α_k^c · A^k). The result is a map at the spatial resolution of the last convolutional layer (typically 7×7 or 14×14 for ImageNet) which we upsample to the input resolution. The lesson covers: (1) the intuition for why the gradient is a meaningful weight of channel "importance" for a class, (2) the role of ReLU in the final formula, (3) compatibility with any CNN (VGG, ResNet, DenseNet, MobileNet), (4) applications to other tasks (image captioning, VQA), (5) differences between GradCAM, GradCAM++ (Chattopadhay et al. 2018) and HiResCAM (Draelos & Carin 2020), (6) typical interpretive pitfalls and sanity checks (Adebayo et al. 2018), (7) typical implementation mistakes.