Calibration and Uncertainty in Machine Learning: Advances, Limits, and Current Challenges

Calibration and uncertainty in machine learning: progress, limits, and current challenges.

If the current state of calibration and uncertainty quantification in machine learning had to be summarized in a single idea, it would be this: we are now much better at measuring model confidence and have reasonably solid tools to improve it in specific scenarios, but we still lack a universal solution that works reliably when data, tasks, or deployment environments change.

Calibration and uncertainty are not the same. Calibration means predicted probabilities match observed frequencies, while uncertainty quantification captures both data noise and model ignorance.

In classical calibration, methods like temperature scaling remain widely used, although their effectiveness depends on validation data matching real-world conditions. Metrics such as ECE are useful but limited.

In uncertainty quantification, approaches like deep ensembles remain highly competitive compared to more theoretically grounded methods. Conformal prediction has gained traction due to its formal guarantees, though these weaken under distribution shifts.

Large language models introduce new challenges: multiple layers of uncertainty and no clear standard for interpreting their probabilities.

Key challenges include distribution shift, proper evaluation, scalability, interpretability, and linking uncertainty to real-world decision-making.

In conclusion, the field has made significant progress, but a core limitation remains: model confidence is easier to estimate in stable environments than in real, changing conditions.