To better understand deep architectures and unsupervised learning, uncluttered by hardware
details, we develop a general autoencoder framework for the comparative study of
autoencoders, including Boolean autoencoders. We derive several results regarding autoencoders
and autoencoder learning, including results on learning complexity, vertical
and horizontal composition, and fundamental connections between critical points and clustering.
Possible implications for the theory of deep architectures are discussed.