Grad-CAM (Gradient-weighted Class Activation Mapping)
XAI method that generates heatmaps showing which image regions a CNN considers most important for its decision.
Grad-CAM visualizes as a heatmap which image regions a CNN uses for its decision – the standard for visual AI explainability.
Explanation
Grad-CAM uses gradients of the last convolutional layer to weight feature maps by relevance. The result is a heatmap showing where the model "looks." Grad-CAM++ improves localization of multiple objects.
Marketing Relevance
Essential for trust in computer vision: medical imaging, autonomous driving, quality inspection – wherever you need to understand what the model sees.
Example
A skin cancer detection model is checked with Grad-CAM: The heatmap shows whether the model actually analyzes the mole or just background artifacts.
Common Pitfalls
Grad-CAM only shows last layer activations – early layer features remain invisible. Heatmaps can be misleading in multi-object scenes. Not suitable for non-CNN architectures.
Origin & History
Selvaraju et al. published Grad-CAM in 2017 (ICCV). Grad-CAM++ (2018) improved multi-object localization. Score-CAM (2020) eliminated gradient dependency. The method is standard in medical AI explainability.
Comparisons & Differences
Grad-CAM (Gradient-weighted Class Activation Mapping) vs. LIME
LIME is model-agnostic and perturbation-based; Grad-CAM uses CNN gradients directly and is faster but CNN-specific.
Grad-CAM (Gradient-weighted Class Activation Mapping) vs. Saliency Map
Saliency maps show pixel-level gradients (noisy); Grad-CAM aggregates at feature map level (smoother heatmaps).