Pertinent negatives

In this section, we will visualize the contrastive explanation for a pertinent negative.

For example, the number 3 could be an 8 with a missing half:

Figure 10.13: 3 is half of 8!

We first display a specific instance:

# @title Generate contrastive explanation with pertinent negative
# Explained instance
idx = 15
X = x_test[idx].reshape((1,) + x_test[idx].shape)
plt.imshow(X.reshape(28, 28));

The output is displayed as follows

Figure 10.14: An instance to explain

We now check the accuracy of the model:

# @title Model prediction
cnn.predict(X).argmax(), cnn.predict(X).max()

We can see that the prediction is correct:

(5, 0.9988611)

We will now create a pertinent negative explanation.

CEM parameters

We will now set the parameters of the CEM explainer.

The Alibi notebook explains each parameter in detail.

I formatted the code in the following text so that each line of code is preceded by the description...