Task Formulation

Saliency methods are widely used in XAI to highlight the regions in an image that most influence a model’s prediction. These methods generate saliency maps, which visually represent the contribution of each pixel to the model’s output. Given an input image I∈ℝch×h×w, where ch, h, and w represent the number of channels, height, and width of the image, a black-box classifier can be described by the function f:ℝch×h×w→ℝC, where C is the number of classes. A saliency method generates a saliency map S=E(I,f)∈ℝh×w, where the saliency map has the same spatial dimensions as the input image.

Saliency methods provide insights into the model’s decision-making by quantifying the relevance of individual pixels in relation to the final output. These methods are evaluated based on their ability to highlight regions that significantly influence the model’s predictions, aligning the model’s rationale with human reasoning.