Formalizing Consistency and Coherence of Representation Learning