Słownik AI
Kompletny słownik sztucznej inteligencji
Platt Scaling
Parametric calibration method using logistic regression to adjust probability scores of predictions.
Isotonic Regression
Non-parametric approach to calibrate probabilities while preserving the monotonic order between predictions and calibrated probabilities.
Temperature Scaling
Simple calibration technique for neural networks applying a single temperature factor to logits before softmax.
Deep Ensembles
Uncertainty estimation method combining multiple models trained with different initializations to capture epistemic uncertainty.
Monte Carlo Dropout
Bayesian approximation technique using dropout during inference to sample predictions and quantify uncertainty.
Bayesian Neural Networks
An approach where the network weights are treated as probability distributions rather than fixed values to model uncertainty.
Conformal Prediction
Theoretical framework guaranteeing probabilistic coverage for predictions through calibrated prediction sets.
Expected Calibration Error
Main metric to evaluate calibration measuring the average difference between confidence and accuracy on prediction bins.
Brier Score
Mesure de la qualité des prédictions probabilistes combinant calibration et refinement, particulièrement utilisée en classification binaire.
Quantile Regression
Méthode de régression estimant les quantiles conditionnels pour construire des intervalles de prédiction asymétriques calibrés.
Gaussian Processes
Non-parametric approach naturally providing uncertainty estimates for predictions via Gaussian distributions.
Variational Inference
Technique d'approximation bayésienne optimisant une distribution de paramètres pour quantifier l'incertitude dans les modèles complexes.