A type of Saliency Map that works by calculating the gradient of the output of a CNN with respect to the input pixels.
Since this method finds a gradient for each pixel, we can highlight each pixel based on the magnitude of their gradient and create a heatmap that shows the most important pixels of the original image.
Determining the gradient of the ReLU can be ambiguous in its non-continuous point, but Vanilla Gradient resolves this ambiguity by letting the gradient be zero when the input to ReLU is zero.
Examples
Saliency map of neural networks performing regression to predict the size of a circle in an image:  Here we see that the pixels around the circle are the ones important for determining its size.