Machine Learning: Future possibilities
- Updated
On this page
Better decisions about when to abstain
Possible future · No delivery date predicted · near-term possibility; no promised delivery date or outcome.
More applications may use calibrated uncertainty to route difficult cases for review. The benefit depends on whether uncertainty remains reliable for the operating population and whether the review process can handle the workload. Abstention is a useful option, not a complete safety argument.
A Gentle Introduction to Conformal Prediction — Angelopoulos and Bates / arXiv · Source date: 2021-07-15 · Checked
Learning with more local control over data
Possible future · Long-term possibility; unresolved · long-term possibility; no promised delivery date or outcome.
Distributed learning could broaden participation without moving every raw record into a central collection. Reaching that goal requires privacy protections, reliable coordination and fair evaluation across participants. The original model-averaging result is a foundation to investigate, not proof that all those challenges are solved.
Communication-Efficient Learning of Deep Networks from Decentralized Data — McMahan and colleagues / arXiv · Source date: 2016-02-17 · Checked