In the previous recipe, Implementing vanilla autoencoders, we implemented the simplest autoencoder. In this recipe, we will build a deep autoencoder using MNIST digits data to demonstrate dimensionality reduction. The data preprocessing will be the same as that of the previous recipe, Implementing vanilla autoencoders. We will extract encoded features (low dimensional) from the encoder network and then use these encoded features in the decoder to reconstruct the original input and evaluate the reconstruction error. We use these encoded features to build a digit classification model.