This sounds like a divide & conquer approach (Sorry if this generalization is too lame). If it can work on less capable device than it will create a new wave of

This sounds like a divide & conquer approach (Sorry if this generalization is too lame). If it can work on less capable device than it will create a new wave of innovations in mobile devices. I am wondering, whether a similar approach is possible with current DL models and will it have any performance improvements over what is existing or whether it will be computationally even more expensive.