Eigen-Cam

Eigen-CAM is a class activation mapping method based on principal component analysis (PCA) of the feature maps — using the first principal component of the activation maps as the saliency map, without requiring class-specific gradients or forward passes.

How Eigen-CAM Works

• Feature Maps: Extract $K$ activation maps from a convolutional layer, each of dimension $H \times W$.
• Reshape: Reshape maps to a $K \times (H \cdot W)$ matrix.
• PCA: Compute the first principal component of this matrix.
• Saliency: Reshape the first principal component back to $H \times W$ — this is the Eigen-CAM.

Why It Matters

• Class-Agnostic: No gradient or target class needed — highlights the most "activated" spatial regions.
• Fast: Just one SVD computation — faster than Score-CAM or Ablation-CAM.
• Limitation: Not class-discriminative — shows what the network attends to, not what distinguishes classes.

Eigen-CAM is the principal attention pattern — using PCA to find the dominant spatial focus of the network without any gradients.