Well, in fairness, it's relatively difficult to do today: first of all, most neural architectures don't fit on mobile, because most of the research is either mo

Well, in fairness, it's relatively difficult to do today: first of all, most neural architectures don't fit on mobile, because most of the research is either more academic than production-focused, or because people are chasing the best possible accuracy as opposed to accuracy under constraints. There's still great applicable research, and in particular the SqueezeNet, QuickNet and MobileNets papers are great in that respect. Then there are some issues with just the tooling. TensorFlow is great, and works great on mobile, but it was a bit hard to get it working. The whole team there is doing an awesome job documenting and improving everything though! It's just early and not enough users have built apps with it to get the kind of production-readiness you get out of non-Machine Learning components.