Grad-CAM Entropy Loss

• The paper presents an auxiliary loss leveraging the Shannon entropy of normalized Grad-CAM maps to promote concentrated and interpretable activation regions.
• It integrates the entropy term with standard cross-entropy loss using a hyperparameter to control the trade-off between classification accuracy and spatial localization.
• Empirical results demonstrate improved Grad-CAM sharpness and contrast on CNN architectures with only a minor accuracy reduction, all without modifying the network structure.

A Grad-CAM-based Dice loss does not appear in the referenced literature. Instead, the relevant approach incorporates an auxiliary loss term based on the Shannon entropy of normalized Grad-CAM maps, aimed at enhancing the interpretability and localization properties of convolutional neural network (CNN) classifiers’ gradient-weighted class activation maps (Grad-CAMs). This entropy-regularized loss combines with the standard classification objective, does not require architectural modifications, and leverages second-order derivatives for optimization. The resulting method enables explicit control over the trade-off between classification accuracy and spatial concentration of explanation maps, facilitating deployment on both deep and embedded architectures without additional network layers (Schöttl, 2020).

1. Formulation of the Grad-CAM Entropy Loss

The technique introduces an auxiliary loss term defined as the spatial Shannon entropy of a normalized Grad-CAM map for the ground-truth class. Let $L_{GC}^{(c)}$ denote the unnormalized Grad-CAM activation map for class $c$:

$$\widetilde M_{ij} = \frac{L_{GC,ij}^{(c)}}{\sum_{k,\ell} L_{GC,k\ell}^{(c)}}$$

$$c_e\bigl(L_{GC}^{(c)}\bigr) = -\sum_{i,j} \widetilde M_{ij}\log\!\left(\widetilde M_{ij}\right)$$

Here, $\widetilde M_{ij}$ is a spatial probability distribution over map locations, ensuring $\sum_{i,j}\widetilde M_{ij}=1$. The term $c_e$ acts as an “interpretability regularizer”, penalizing distributed (high-entropy) attention maps and favoring concentrated regions of activation.

2. Integration With Classification Objective

The total loss function combines standard cross-entropy on softmax outputs with the entropy term scaled by a new non-negative hyperparameter $\beta$:

$$\mathcal{L} = \mathcal{L}_{CE}\bigl(y_{\rm true}, y_{\rm pred}\bigr) + \beta\,c_e\left(L_{GC}^{(c)}\right)$$

where $$\mathcal{L}_{CE}(y_{\rm true},y_{\rm pred}) = -\sum_{m=1}^C y_{\rm true}^m \log y_{\rm pred}^m$$. The hyperparameter $\beta$ governs the trade-off, with larger values enforcing greater localization in Grad-CAM maps at potential expense to classification accuracy. For $\beta=0$, the scheme reverts to conventional training.

3. Grad-CAM Map Computation and Second-Order Gradients

During training, the raw Grad-CAM map for class $c$ is constructed as follows. Let $A^k \in \mathbb{R}^{H\times W}$ denote the $k$-th feature map from the final convolutional layer, and $y^c$ the pre-softmax logit for class $c$:

$$\alpha_k = \frac{1}{Z}\sum_{i,j}\frac{\partial y^c}{\partial A^k_{ij}}, \quad L_{GC,ij}^{(c)} = \mathrm{ReLU}\left(\sum_k \alpha_k A^k_{ij}\right)$$

where $Z = H \cdot W$.

Since the loss involves entropy of the normalized Grad-CAM map, and $L_{GC}^{(c)}$ in turn depends on gradients of the logit with respect to feature maps, the backward pass computes derivatives of derivatives (second-order gradients) with respect to the model parameters. In modern frameworks supporting higher-order differentiation, this is automatically handled, at the cost of increased computational overhead.

4. Definitions and Symbol Table

Symbol/Term	Meaning
$x$	Input image, $$x\in\mathbb{R}^{H_{\rm in}\times W_{\rm in}\times C_{\rm in}}$$
$A^k$	$k$-th feature map in last conv layer, $\in\mathbb{R}^{H\times W}$
$y^c$	Logit (pre-softmax score) for class $c$
$y_{\rm pred}$	Prediction vector, $\mathrm{softmax}(y)$
$y_{\rm true}$	One-hot ground truth
$\mathcal{L}_{CE}$	Cross-entropy loss
$\alpha_k$	Grad-CAM scalar, global-average gradient over feature map $A^k$
$L_{GC,ij}^{(c)}$	Grad-CAM activation map (raw, unnormalized) for class $c$
$\widetilde M_{ij}$	Normalized Grad-CAM activation map
$c_e(L_{GC}^{(c)})$	Entropy of normalized Grad-CAM map
$\beta$	Weight for Grad-CAM entropy term

5. Training Procedure and Workflow

A condensed outline of the training loop integrating the Grad-CAM entropy term:

1. Initialize network parameters $\theta$; select $\beta \ge 0$.
2. For each epoch and mini-batch:
   • Perform forward pass to compute feature maps $A^k$, logits $y$, and predictions.
   • Compute standard cross-entropy loss.
   • Retain gradients of logits with respect to feature maps.
   • Calculate Grad-CAM weights and construct the activation map.
   • Normalize $L_{GC,ij}$ to obtain $\widetilde M_{ij}$.
   • Evaluate the entropy $c_e$.
   • Form the combined loss $\mathcal{L}_{CE} + \beta c_e$.
   • Backpropagate (including required second-order gradients) and update parameters.

The key computational cost arises in the backward pass, where second-order derivatives are necessary due to the dependency of $c_e$ on $L_{GC}^{(c)}$, which itself is a function of gradients.

6. Empirical Findings on Interpretability and Classification

Experiments conducted on a ResNet-50 backbone with additional dense layers, using PASCAL VOC 2012 object crops, illustrate the quantitative effect of the entropy penalty:

• With $\beta=0$ (no penalty), test accuracy stabilized at ≈0.94 with relatively high-entropy, diffuse Grad-CAM maps.
• For $\beta=100$, test accuracy decreased marginally to ≈0.92; however, CAM entropy dropped from ≈0.99 to ≈0.92, “ellipsoidal area” ca contracted by ≈20%, and “dispersion” cd increased six-fold, indicating more decisive, sharply focused attention maps.
• In a class “dog” example, Grad-CAM contrast increased by 114%, highlighting significantly improved localization on relevant regions.
• A trade-off was observed in training speed: higher $\beta$ values approximately doubled backward pass duration due to second-order gradient tracing.

7. Applicability and Limitations

The approach can be incorporated into any standard CNN classifier without structural modification and generalizes across deep and embedded platforms. The hyperparameter $\beta$ permits continuous adjustment between interpretability and prediction performance, with only minor degradation in classification accuracy for substantial gains in map localization and dispersion. There is no introduction of a Dice-style overlap loss; instead, the method targets reduction of spatial entropy in the Grad-CAM explanation itself. This suggests that entropy penalization is a lightweight and effective means to encourage more interpretable, user-tractable explanations from standard CNN-based image classifiers (Schöttl, 2020).