The author here used the MNIST set of handwritten numbers, but that seems an unnecessary constraint. An infinite (practically) training set is available in the form of existing fonts, which can be used to generate a training set. A set of distortions can be added in automatically to give the net a little more information to work with, and then this same methodology can be applied to generate the single glyph ambigrams.
bigram<->character ambigrams are also feasible, but the search space gets pretty big.