Grad-CAM (Gradient-weighted Class Activation Mapping) is a visual explanation technique that produces a coarse localization map highlighting the important regions in an image — using the gradients flowing into the last convolutional layer to weight the activation maps by their importance for the target class.

How Grad-CAM Works

• Gradients: Compute gradients of the target class score with respect to feature maps of the last conv layer.
• Weights: Global average pool the gradients to get importance weights $alpha_k$ for each feature map $k$.
• CAM: $L_{Grad-CAM} = ReLU(sum_k alpha_k A_k)$ — weighted sum of feature maps, ReLU keeps only positive influence.
• Upsampling: Upsample the CAM to input image resolution for overlay visualization.

Why It Matters

• Model-Agnostic: Works with any CNN architecture that has convolutional layers.
• Class-Discriminative: Different target classes produce different heat maps — shows what the model looks for per class.
• No Retraining: Post-hoc technique — no modification to the model architecture or training.

Grad-CAM is seeing what the CNN sees — highlighting the image regions that most influenced the classification decision.