Computations are computations; you certainly don't need to use it for deep learning, but you might want to introduce some new graph-level optimizations if they are appropriate to what you are doing and haven't shown up in deep learning.