← AI Terminology

Calibration

Calibration is the degree to which a model's predicted probabilities match actual outcome frequencies — a perfectly calibrated model that says "70% chance of rain" should be right 70% of the time it says that.

It measures whether probability outputs are trustworthy, not just directionally correct.
Why It Matters in AI
High accuracy does not imply good calibration. A model that predicts 0.99 probability for every positive and is correct 80% of the time is wildly overconfident. In medical diagnosis, fraud detection, and financial risk, decision-makers use the probability directly — not just the label — so calibration errors propagate into bad decisions. Calibration is also critical for LLM uncertainty: a model that confidently states falsehoods is poorly calibrated.
Key Points
Aspect Description
ECE Expected Calibration Error: weighted average of reliability curve deviation from diagonal
Platt scaling Fit a logistic regression on validation set outputs to recalibrate probabilities
Overconfidence Most neural networks are overconfident — predicted probability > actual frequency
LLM calibration LLMs often confidently hallucinate — calibration alignment is an active safety research area
Reliability curve Plot predicted probability (x) vs observed frequency (y) — perfect calibration is the diagonal
Temperature scaling Post-hoc calibration: divide logits by T > 1 to soften probabilities without retraining
Simple Analogy
A confident doctor who says "I'm 95% sure it's benign" should be right 95 times out of 100. If they're only right 70 times, they're overconfident — their probability numbers can't be taken at face value. Calibration is the check on whether confidence scores actually track reality.
Common Usage Examples
  • sklearn.calibration.calibration_curve(y_true, y_prob) — plot the reliability diagram
  • CalibratedClassifierCV(base_estimator, method='isotonic') — isotonic regression recalibration
  • Temperature scaling: logits_calibrated = logits / T where T found on validation set
  • sklearn.metrics.brier_score_loss and log_loss — both sensitive to calibration
  • DeepMind calibration paper (Guo et al., 2017): modern neural networks are badly overconfident
Summary
In short: Calibration is whether a model's confidence scores mean what they say — a model that says "80% confident" should be right 80% of the time, not 95% or 60%.