In praise of the autoencoder