Confusion matrices are great to check which classes get mixed up together, but then you're left to plot, or visualize the wrong examples yourself in order to figure out why said classes are confusing your model.
With Facets the examples can be displayed directly inside of the confusion matrix as shown on Google's blog with the CIFAR dataset:
And you can zoom in on the different classes and better visualize where the problem might lie:

This repository provides a more thorough example of using Facets to create such a matrix.
To do so, a small CNN is trained on the MNIST dataset using Keras and Tensorflow and a custom function is used to transform the image array into a "sprite atlas" as used by Facets Dive.