Yes I understand that conceptually. But the middle autoencoder layer is not really an output layer, what would constrain it to linear representations? I assume

Yes I understand that conceptually. But the middle autoencoder layer is not really an output layer, what would constrain it to linear representations? I assume normal output layers are constrained to linear representations by the supervised training process.