Autoencoders do not have that many tangible parameters to tune and retain information when feature representation is done in some different space in a layer.
I have 5 layers in keras , and have got 50% accuracy on Autoencoder.
How can I make feature representation more preserved in Autoencoder?