In the next sections of the course, we will discuss many methods for producing saliency maps. These can provide great
Axioms are desirable mathematical properties of XAI methods. As we discuss the 7 common axioms in Figure 1, you will
Much of Explainable AI (XAI) research revolves around developing new methods to explain model predictions. What is equally important is
LIME uses perturbations of an individual instance to build a model around that instance. We call this a surrogate model.
When it comes to saliency maps, Integrated Gradients (IG) is a strong, well-justified choice with solid theoretical foundations. It could
With DeepLIFT, we are interested in explaining the difference between the prediction on an input image and a baseline image
Until now, we’ve focused on methods that use the gradients of an input image (\(\frac{\partial y^c}{\partial X}\)). These methods all
Grad-CAM and Guided Backpropagation (GBP) are both simple and effective methods for producing saliency maps. If you had to make
Vanilla gradients have many limitations. When it comes to reliability, the key issue is that it produces noisy saliency maps.
In deep learning, gradients are typically used to train a model. They update parameters and ultimately form the complex features