Grad-CAM and Visual Saliency Maps

A classifier produces a score for a chosen class. Grad-CAM, introduced by Selvaraju and colleagues, backpropagates that target score’s gradients into a convolutional feature layer, weights its channels and combines them into a spatial map. Upscaling and overlaying the map on the image produces familiar warm-colored highlights. The map is class-specific: asking about one output can yield a different picture than asking about another. Its spatial detail is limited by the feature layer, so it is usually coarser than a segmentation mask. The visual is useful for generating questions. If the model predicts “horse” while lighting up a stable logo or background, the team has a possible shortcut to investigate. If it highlights part of the animal, that may be reassuring but still does not prove the model used the same evidence a human would. Map appearance depends on the chosen layer, target score, normalization and rendering. A red patch is not a probability that each pixel belongs to the class and does not show every interaction in the network. Testing requires interventions. Crop or mask a suspected cue carefully, change backgrounds while keeping the target, and compare predictions with a suitable control. An intervention can itself alter the image distribution, so one changed score is not always definitive. For localization claims, compare against independent region annotations and report how well the heat map overlaps relevant structures. In a high-stakes setting, experts should evaluate model outputs and failure cases rather than relying on a colorful explanation to certify a decision. Grad-CAM applies to architectures with suitable differentiable spatial features; implementations differ. Other saliency methods may emphasize gradients at input pixels or use perturbations, so their maps should not be conflated. Use the visualization to direct auditing, document how it was produced and verify the behavior it suggests. A convincing heat map cannot compensate for weak external validation or incorrect labels.