How so? You still aren't saying what the benefits are. OK, to be fair I said tradeoffs, but I'm really wondering what the point of this work is. I mean my models are ridiculously tiny already, even when considered for use on mobile devices, so clearly that's not the win here. What is?
Your models might be ridiculously tiny, but a lot of people's models are not. Take a loot at any research paper in vision, speech or language. The models are gigantic.
