<aside> 🧑‍🏫 Relevance-CAM outperforms other CAM-based methods in recognition and localization evaluation in layers of any depth.

</aside>

Background

• CAM
• Grad-CAM

• Layer-wise Relevance Propagation (LRP)
• Contrastive Layer-wise Relevance PRopagation (CLRP)

Gradient Issue

Noisiness and discontinuity

• Shattered gradient problem

→ ReLU/Sigmoid activations in deep networks

• Piece-wise linearity of gradients

→ relations to next pixels are lost

Explanation to sensitivity

Grad-CAM : importance of an activation map is measured by gradient of output w.r.t the activation map → does not take account for activation values(only sensitivity)

what we want to explain is how much an activation map contributes to a target class output and not how sensitive an activation map is. This issue is called False Confidence

Relevance-weighted Class Activation Map