Coding Exercise on Class Activation Maps

Ablation CAM: Gradient-free approach to GradCAM

Grad-CAM suffers from the problem of gradient saturation, which causes the backpropagated gradients to diminish and, therefore, adversely affects the quality of visualizations. Ablation-CAM is a gradient-free approach to GradCAM that avoids using gradients and, at the same time, produces high-quality CAMs.

Overview of GradCAM

Let’s assume that $\mathcal{F}_l \in \R^{h \times w}$ is the $l^{th}$ feature map obtained from the penultimate convolutional layer ($h$ and $w$ are the height and width of the feature map) and $\mathcal{F}_l(i,j)$ represents the activation of the $(i,j)^{th}$ neuron in the $l^{th}$ feature map (total $L$ feature maps in $\mathcal{F}$).

GradCAM computes the gradient $\mathcal{G}_l(i,j) = \nabla_{\mathcal{F_l(i,j)}} f^{k^*}(X)$ of the prediction score $f^{k^*}(X)$ with respect to a feature $\mathcal{F}_l(i,j)$. The gradient $\mathcal{G}_l(i,j) \in \R^{h \times w}$ is then average pooled across height and width to be used as weights in the CAM equation: