AI 용어집
인공지능 완전 사전
Expected Calibration Error (ECE)
A weighted metric calculating the average calibration error by dividing predictions into confidence bins and measuring the difference between confidence and accuracy.
Maximum Calibration Error (MCE)
A metric identifying the worst-case disagreement between confidence and accuracy across all confidence bins, useful for assessing extreme risks.
Reliability Diagram
A graphical visualization comparing predicted probabilities (confidence) to empirical frequencies (accuracy) to visually assess a model's calibration.
Histogram Binning
A calibration method that discretizes prediction scores into bins and replaces each score with the empirical frequency of its corresponding class.
Beta Calibration
A parametric technique using the Beta regression function to model the relationship between raw scores and calibrated probabilities, suited for binary predictions.
Dirichlet Calibration
A multi-class extension of Beta calibration using the Dirichlet distribution to simultaneously calibrate all classes with interdependencies.
Logistic Calibration
A family of parametric methods based on logistic regression to adjust predicted probabilities, including Platt Scaling as a special case.
Class-wise Calibration
Calibration performed separately for each class in a multi-class problem, as opposed to global calibration which considers all classes simultaneously.
Top-label Calibration
Calibration specific to the prediction with the highest probability, particularly important in systems where only the best prediction is used.
Posterior Calibration
Process of adjusting posterior probabilities to match the true conditional distributions given the input features.
Adaptive Calibration
Calibration methods that dynamically adapt to changing data distributions, continuously readjusting calibration parameters.
Multi-class Calibration
Extension of binary calibration techniques to multi-class problems, requiring simultaneous calibration of probability distributions over multiple classes.