ablation cam

**Ablation-CAM** is a **class activation mapping variant that determines feature map importance by ablation** — systematically removing (zeroing out) each feature map and measuring the drop in the target class score, providing a principled, gradient-free importance measure. **How Ablation-CAM Works** - **Baseline**: Record the target class score with all feature maps present. - **Ablation**: For each feature map $A_k$, zero it out and re-forward — record the score drop $Delta s_k$. - **Weights**: The importance weight for map $k$ is proportional to the score drop when $A_k$ is removed. - **CAM**: $L_{Ablation} = ReLU(sum_k Delta s_k cdot A_k)$ — weight maps by their ablation importance. **Why It Matters** - **Causal**: Ablation directly measures causal importance — "removing this feature reduced the score by X." - **No Gradients**: Like Score-CAM, avoids gradient issues — suitable for non-differentiable models. - **Validation**: Can validate Grad-CAM explanations by checking if gradient-based and ablation-based importance agree. **Ablation-CAM** is **remove-and-measure** — determining each feature map's importance by testing what happens when it's removed.