Chapter 8: Visualizing Convolutional Neural Networks

Up to this point, we have only dealt with tabular data and, briefly, text data in Chapter 6, Local Model-Agnostic Interpretation Methods. This chapter will exclusively explore interpretation methods that work with images and, in particular, with the Convolutional Neural Network (CNN) models that train image classifiers. Typically, deep learning models are regarded as the epitome of black box models. However, one of the benefits of a CNN is how easily it lends itself to visualization, so we can not only visualize outcomes, but every step of the learning process with activations. The possibility of interpreting these steps is rare among so-called black box models. Once we have grasped how the CNN is learning, we will study how to use state-of-the-art gradient-based attribution methods such as Saliency Maps and Grad-CAM to debug class attribution. Lastly, we will extend our attribution debugging know-how with perturbation-based attribution...