Smooth Gradient Saliency

Smooth gradient saliency

The vanilla gradient saliency computes the saliency map using the vanilla gradient $G$ of the score function $f^{k^*}(X)$ with respect to input $X$. However, these gradients are often noisy, highlighting some pixels that seem randomly selected for the human eye. Such unwanted visual noise in the saliency maps can alter human judgment sometimes. For example, if the saliency map is noisy and highlights random background pixels, a data scientist might think the background information is relevant to the prediction, which should not be the case.

Smooth gradient (SmoothGrad) saliency reduces visual noise and produces sharper saliency maps. The main idea is to smoothen the gradients (used for computing the saliency map) by averaging the vanilla gradients of multiple noisy versions of the input image. These smoothened gradients have a denoising effect that helps reduce the visual noise in the saliency maps and makes them sharper. The figure below illustrates this concept.