I tried to model my data, and for now have a decent model using SVM. I thought that maybe using an Autoencoder for creating some nontrivial representations of data could help.
My problem is I have only seen Autoencoders with binary inputs/outputs and logistic neurons in hidden layer. Does making an Autoencoder with logistic mapping to hidden units and then linear mapping to output unit makes sense? Does anybody have experience with such models and can say if they're worth a shot? Do you know any papers about such models or about learning interesting representations from continuous inputs? (I have already tried PCA).
Thank you in advance.