Understanding regularization in autoencoders