On the face of it, autoencoders seem to be literally the simplest class of deep learning models.