Machine Learning: Research directions

Conformal methods motivate work on uncertainty for structured outputs, time series and distribution shift. Useful guarantees require careful assumptions and calibration. Making a model express uncertainty does not automatically make its uncertainty valid after the environment changes.

Federated learning addresses the tension between locally held data and shared models. Unequal device participation, differing data and communication constraints complicate evaluation. It is a design space with workload-specific tradeoffs, rather than a drop-in replacement for every centralized training system.