Based on what I've read and seen so far, one advantage and use of autoencoder neural networks, is to find/cherry-pick your useful features. That's if the Autoencoder is trained well, then the inner layer containing the compressed data (that's latent layer, or bottleneck) has very valuable data in it. This layer can be used as input to another network and make the other network training much much easier/faster/accurate.
In another word, we get rid of the decoder part, and attach new layers to the encoder part.
Questions: 1.Can/should the new layers have different activation functions?
Is there possibly any valid reason for a model to add more node count per layer, after the latent layer? (I mean when using the latent as the input to a new network, does it ever make sense to widen the network again?)
How one does this procedure using keras? I know how to train a model, and an autoencoder, but how exactly I combine them?
I Have the Autoencoder trained, and know the form of the desired output for the new network. And I only know how to make networks using sequential method in keras and tensorflow.