Production ML systems already do this: servers used for peak capacity are repurposed during lulls to retrain the model and do tasks like compact data. That retrained model is then shared across the system.
So I’d argue that ML systems already “sleep”, it just looks more like dolphin sleep due to their similar need for continuous operation.
So I’d argue that ML systems already “sleep”, it just looks more like dolphin sleep due to their similar need for continuous operation.