In some applications you may want to run a continual learning algorithm so it will continually train on new data as well as make inferences.
I always wonder how one would guarantee stability / convergence when learning in the field. Sounds great, but how can this be pulled off in a reliable way? Just dragging down the learning rate does not suffice in my oppinion, as for effective training you should be at the boundary of stability and greedy updates from my experience
My main computer is a laptop with a Intel UHD Graphics 620 card :) Having a small card like this to speed up training would be so cool.
To fine tune the model to the specific customer needs. For example, learning the parameters of they user's face.