I have been reading lately on autoencoders a lot. I just wanted to summarize my understanding of denoising autoencoders. As far as I understand they can be
Fully connected (in which case, they will be over-complete autoencoders)
The reason I say it should be over-complete is that the objective is to learn new features and I think extra neurons in the latent layer would help. There is no reason to have a lesser number of neurons because compressing is not the objective. I just want to understand is this the right thinking.