If it was possible and without any performance costs, would it be good to be able to modify training parameters in real-time while the algo keeps running and instantly affected by any param change?
Meaning zero latency to always access the latest params locally inside the algorithm, as if we could remotely modify any internal variable in the training memory.
If this is technically possible, would this be useful for AIOPS/MLOPS ?