Score-Cam

Score-CAM is a gradient-free class activation mapping method that weights activation maps by their contribution to the model's confidence — replacing gradient-based weighting with perturbation-based importance, avoiding issues with noisy or vanishing gradients.

How Score-CAM Works

• Activation Maps: Extract feature maps from the target convolutional layer.
• Masking: For each feature map, normalize and use it as a mask on the input image.
• Scoring: Feed each masked image through the model to get the target class score (the "importance" of that map).
• Combination: $L_{Score-CAM} = ReLU(\sum_k s_k \cdot A_k)$ — weight maps by their confidence scores.

Why It Matters

• No Gradients: Avoids gradient noise and saturation issues — more stable explanations.
• Faithful: Importance weights directly measure each map's effect on the model's confidence.
• Trade-Off: Requires $N$ forward passes (one per activation map) — slower than Grad-CAM but more robust.

Score-CAM is measuring importance by masking — directly testing each feature map's effect on the prediction for gradient-free visual explanations.