AI 용어집
인공지능 완전 사전
Multiclass Brier Score
Calibration evaluation metric measuring the mean squared error between predicted probabilities and ground truth indicators across all classes.
Top-K Calibration
Extension of calibration evaluating the agreement between the probability that the true class appears in the top K confident predictions and its empirical frequency.
Multiclass Platt Scaling
Generalization of binary logistic regression to the multiclass case through one-vs-all or one-vs-one approaches to individually calibrate each class.
Multiclass Histogram Binning
Non-parametric method partitioning the probability space into intervals and adjusting predictions by the empirical frequency observed in each bin for all classes.
Pairwise Calibration
Strategy calibrating relative probabilities between each pair of classes to ensure comparative consistency in the multiclass space.
Multiclass Reliability Diagram
Calibration visualization comparing mean confidence to mean accuracy per bin to graphically assess the adequacy of multiclass predictions.
Multiclass Isotonic Regression
Extension of isotonic regression to the multiclass case through approaches like PAVA (Pool Adjacent Violators Algorithm) or multi-output variants.
Bayesian Binning Quantile
Calibration method optimally determining bin boundaries using Bayesian principles to minimize multiclass calibration error.
Adaptive Calibration Error (ACE)
Variant of ECE using adaptive size bins based on the prediction distribution to reduce the variance of the calibration estimator.