Good thread on the last mile problem for deep learning. The main thing that’s worked in the successful production ML apps I’ve supervised (from variant calling to anti-fraud) is a cheap, fast feedback loop to turn new in-field training examples into online model updates.