KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Platt Scaling
Parametric calibration technique using logistic regression on the model's scores to adjust the predicted probabilities.
Isotonic Regression
Non-parametric calibration method that fits a piecewise-constant monotonic function to calibrate predictions.
Temperature Scaling
Technique simple de calibration qui ajuste un paramètre de température dans la fonction softmax pour améliorer la calibration des réseaux de neurones.
Brier Score
Evaluation metric that measures the mean of the squared differences between predicted probabilities and actual outcomes to assess calibration.
Reliability Diagram
A visualization tool that compares predicted probabilities to observed frequencies in bins to diagnose calibration.
Expected Calibration Error (ECE)
Quantitative measure that calculates the average calibration error weighted by the distribution of model predictions.
Bayesian Calibration
Bayesian approach that uses prior distributions to estimate uncertainty and calibrate probabilistic predictions.
Venn-Abers Calibration
Probabilistic calibration method that guarantees valid predictions even under weak assumptions about the data distribution.
Conformal Prediction
Statistical framework that provides prediction sets with coverage guarantee, complementary to probability calibration.
Multi-class Calibration
Extension of calibration techniques to the multi-class case where alignment must be maintained across all classes.
Calibration for Imbalanced Data
Specialized techniques for calibrating models on imbalanced datasets where class frequencies vary considerably.
Post-hoc Calibration
Methods applied after the complete training of the model to adjust probabilities without modifying prediction performance.