Some Notes on the mathematics of Toy Autoencoding Problems — LessWrong