Saliency Map in AI Explainer
A saliency map is a visual attribution that highlights parts of an input associated with a model output. In computer vision, it often appears as a heat map over pixels; related displays can highlight words, time steps, or sensor regions.
A basic gradient saliency map measures how sensitive the selected output is to a small change in each input value. Other methods smooth gradients, modify backpropagation, or perturb regions to produce a clearer visualization.
Saliency maps are useful for quick inspection, but they can be noisy, unstable, or similar across different model outputs. A bright area means the method assigned importance under its calculation, not that the region caused the real-world event.
Good practice compares several examples, tests randomization and perturbation, records preprocessing, and avoids cherry-picking attractive maps. Explanations should be paired with quantitative evaluation and domain review.
A saliency map highlights input regions whose small changes have a strong local effect on a model's selected output.
Key terms include saliency, heat map, gradient, sensitivity, pixel, perturbation, localization, stability and sanity check.
Many saliency methods ask which small input changes affect the chosen score most.
The resulting values are displayed over the original input.
Visual clarity does not guarantee faithfulness, stability, or causal meaning.
Randomization tests and controlled perturbations help assess the explanation.
Explore more "Explainers"
Discover additional explainers across politics, science, business, technology, and other fields. Each explainer breaks down a complex idea into clear, everyday language—helping you better understand how major concepts, systems, and debates shape the world around us.