Saliency Map
Visualization showing which input pixels or tokens have the greatest influence on model output, based on gradients.
Saliency maps show via gradients which input regions influence a model most – fast but often noisy without SmoothGrad or Integrated Gradients.
Explanation
Saliency maps compute the gradient of the output with respect to the input. High gradient values indicate sensitive regions. Variants: Vanilla gradients, SmoothGrad, Integrated Gradients, DeepLIFT.
Marketing Relevance
Fastest form of visual explainability for deep learning. Useful for debugging but often shows noisy or misleading results without refinement.
Common Pitfalls
Vanilla gradients are often noisy. Saliency maps can focus on irrelevant features. Hard to validate without ground truth.
Origin & History
Simonyan et al. introduced saliency maps for CNNs in 2013. SmoothGrad (2017) reduced noise through averaging. Integrated Gradients (Sundararajan et al., 2017) solved theoretical issues. Now a standard debugging tool in computer vision.
Comparisons & Differences
Saliency Map vs. Grad-CAM
Saliency maps work at pixel level (noisy); Grad-CAM at feature map level (smoother, more interpretable).
Saliency Map vs. SHAP
Saliency maps use gradients (fast but approximate); SHAP uses Shapley values (slower but theoretically grounded).