Neural networks, deep learning, NLP · insight-1416
Re-check calibration on shifted data, because a temperature fitted in distribution may not travel
- Evidence type
- empirical finding
- Difficulty
- hard
- Roles
- ML engineer, Data scientist
- Industries
- cross-industry
Summary: situation, action, reason, boundary
Situation
A temperature-scaled network is deployed to a new region, season or device mix and its probabilities feed a threshold chosen on the old validation split.
Action
Keep a small labelled sample from each new segment, recompute reliability and expected cost there, and recalibrate per segment or add a shift alarm on the confidence distribution.
Reason
Calibration methods are fitted to one distribution. A benchmark under dataset shift found that the quality of predictive uncertainty degrades as the shift grows and that post-hoc recalibration often falls short.
Boundary
Any deployed neural classifier whose probabilities enter a threshold or an expected-cost rule.
Symptom or risk
Confidence stays high while accuracy falls, so the fixed threshold admits more wrong predictions than the calibration plot from the original split suggested.
Exceptions
Stable, closed-world inputs with monitored drift.
Trade-offs
Needs fresh labels, which arrive late in many domains; use delayed audits.
Pitfalls
Monitoring accuracy only; a calibration failure shows up in cost before accuracy moves.
Lessons and concepts
Claims and their sources
Guo et al. report that modern neural networks are often overconfident and that temperature scaling is a simple effective post-hoc remedy on in-distribution data.
Ovadia et al. report that predictive uncertainty degrades under dataset shift and that post-hoc recalibration often falls short.