Parameter Choice and Neuro-Symbolic Approaches for Deep Domain-Invariant Learning