Grad-CAM and Variants

Grad-CAM trades pixel precision for a more readable answer: which spatial regions in a convolutional feature map support this class?

For most nontechnical stakeholders, a regional heatmap is easier to interpret than a field of high-frequency gradients.

ℹ

Class-weighted feature maps

Grad-CAM averages class gradients over spatial positions to get channel weights $\alpha_k^c$ , then combines feature maps $A^k$ :

$L^c=\operatorname{ReLU}\left(\sum_k\alpha_k^c A^k\right)$

ReLU keeps evidence that supports the selected class.

Grad-CAM, Step by Step

Step 1 of 8

Forward-propagate the image

Run the input image through the full CNN. Every convolutional layer computes its activations until the last fully connected layer produces raw class scores (logits).

Formula

αᵏ = (1/Z) Σᵢⱼ ∂yᶜ/∂Aᵏᵢⱼ

L^c = ReLU(Σₖ αᵏ Aᵏ)

Aᵏ = feature map k, yᶜ = class score before softmax

Forward pass → class score

→

CNN layerslogits

house

cat

car

tree

Gradients at last conv layer

roof shape

α=0.72supports

vertical edges

α=0.48supports

sky texture

α=-0.31suppresses

αᵏ × feature map → weighted sum

α=0.72

α=0.48

α=-0.31

→

Σ (pre-ReLU)

ReLU → normalized heatmap

ReLU + normalize

→

Upscaled overlay on input

Class score (house)

—

Channels weighted

—

Negative contributions

—

Step through the full Grad-CAM pipeline. Each stage unlocks the next, so you can follow the gradient from raw class score backward to the spatial heatmap overlaid on the original image.

Layer choice matters. Late layers carry stronger semantics and lower spatial resolution. Earlier layers retain detail and may highlight generic edges.

Variants such as HiResCAM and LayerCAM change how relevance is aggregated: HiResCAM multiplies gradients and activations element-wise before summing, recovering finer spatial detail that Grad-CAM's global average discards; LayerCAM applies the same idea per-pixel and can then combine maps across multiple layers, trading some class-discriminability for higher resolution.

Compare variants on correct and incorrect predictions, then evaluate them with deletion, insertion, or domain-expert review.

Checkpoint

Why does Grad-CAM usually use a late convolutional layer?

←PreviousSaliency MapsAttribution-Based Explanations Next→Integrated GradientsAttribution-Based Explanations