R Deep Learning Essentials - Second Edition

Accessing the features of the auto-encoder model

We can extract the deep features from the model, that is, the values for the hidden neurons in the model. For this, we will use the model with 16 hidden nodes. We will examine the distribution of correlations using the ggplot2 package, as shown in the following code. The results are shown in Figure 9.5. The deep features have small correlations, that is, usually with an absolute value of <.20. This is what we expect in order for the auto-encoder to work. This means that the features should not duplicate information between them:

encoder <- keras_model(inputs=input_layer, outputs=encoder)encodings <- encoder %>% predict(test.x)encodings<-as.data.frame(encodings)M <- cor(encodings)corrplot(M, ...