Breaking down the autoencoder