Grad-CAM

Grad-CAM computes a coarse localization map by globally averaging the gradients of a target class score with respect to the activation maps of a chosen convolutional layer. These averaged gradients serve as neuron importance weights, which are then used to compute a weighted linear combination of the forward activation maps. The result is a heatmap that highlights the discriminative image regions used by the CNN for a specific prediction. Unlike purely gradient-based methods, this weighting mechanism captures the importance of entire feature maps, not just individual pixels.

A key advantage of Grad-CAM is its applicability to any CNN architecture without requiring architectural modifications or retraining. It works with:

• Fully-convolutional networks for semantic segmentation
• CNNs with global average pooling (like ResNet)
• CNNs with fully-connected layers (like VGG) This flexibility stems from its reliance on gradient flow, which is a universal property of differentiable networks. The technique only requires access to the final convolutional layer's feature maps and the gradient of the target output with respect to those maps.

Standard Grad-CAM produces low-resolution heatmaps due to the spatial downsampling in deep convolutional layers. To overcome this, Guided Grad-CAM fuses the class-discriminative localization of Grad-CAM with the fine-grained pixel-space detail of Guided Backpropagation. This is achieved by performing an element-wise multiplication of the upsampled Grad-CAM heatmap with the Guided Backpropagation saliency map. The result is a high-resolution, class-specific visualization that shows both where the model is looking and which precise pixels within that region are most influential.

Grad-CAM can be extended to generate counterfactual explanations by negating the gradient of the target class. This produces a heatmap highlighting the regions that, if removed or altered, would increase the model's confidence in the target class. This is particularly useful for:

• Failure mode analysis: Understanding why a model predicted incorrectly
• Model debugging: Identifying spurious correlations the model relies on
• Adversarial robustness: Visualizing the minimal regions needed to flip a prediction

The Grad-CAM principle extends beyond image classification. Grad-CAM++ improves localization for multi-instance images by using a weighted average of positive partial derivatives for better object coverage. For vision-language tasks, Textual Grad-CAM uses the gradient of a matching score between image regions and text queries to localize semantic concepts. In video understanding, Grad-CAM is applied to 3D CNNs by treating spatio-temporal activation maps, highlighting the frames and regions critical for action recognition.

The faithfulness of Grad-CAM is evaluated using perturbation-based metrics. By removing the most salient pixels identified by the heatmap and measuring the drop in class confidence, one can quantify how accurately the explanation reflects the model's decision process. Key metrics include:

• Pointing Game: Measures if the maximum heatmap value falls within a ground-truth bounding box
• Insertion/Deletion Curves: Tracks the change in probability as pixels are progressively added or removed in order of attribution rank
• Energy-Based Pointing Game: A more robust variant that considers the fraction of total heatmap energy inside the object bounding box

A gradient-free class activation mapping method that eliminates the reliance on potentially noisy gradients. Instead of using gradient flow, Score-CAM weights each activation map by its increase in confidence score when applied as a mask to the input. This produces sharper, more faithful localization maps without the shattered gradient problem that plagues gradient-based methods.

The foundational visualization technique that computes the gradient of the class score with respect to input pixels. Unlike Grad-CAM's coarse, class-discriminative heatmaps, raw saliency maps highlight every pixel influencing the prediction, often resulting in visually noisy outputs. Key limitations include:

• Susceptibility to shattered gradients
• No class-specific localization structure
• High sensitivity to small input perturbations

A modified backpropagation algorithm that sharpens feature visualizations by restricting gradient flow to only positive gradients and positive activations. This suppresses negative contributions that can cancel out important features. When combined with Grad-CAM via pointwise multiplication, it produces the high-resolution Guided Grad-CAM visualization that has both class-discriminative localization and fine-grained detail.

An axiomatic attribution method that satisfies the completeness and sensitivity-n axioms. It computes feature importance by accumulating gradients along a straight-line path from a baseline input (typically a black image) to the actual input. Unlike Grad-CAM's coarse localization, Integrated Gradients provides pixel-level attribution with mathematical guarantees, making it suitable for regulatory compliance scenarios.

A decomposition technique that redistributes a model's prediction score backwards through the network using a conservation property. Each neuron receives relevance from higher layers and redistributes it to lower layers. Key distinctions from Grad-CAM:

• Operates on individual predictions, not class-specific maps
• Requires no gradients
• Guarantees total relevance equals the output score

A technique that sharpens gradient-based saliency maps by averaging gradients from multiple noisy copies of the same input. By adding Gaussian noise and computing the mean gradient, SmoothGrad suppresses the high-frequency noise that causes shattered gradient artifacts. This method can be applied as a preprocessing step to improve the visual coherence of Grad-CAM's underlying gradient signal.