vix.ing · top · new · best · stats

Beyond temperature scaling: Obtaining well-calibrated multiclass probabilities with Dirichlet calibration

2019/10/28 by Meelis Kull, Miquel Perelló-Nieto, Kull, Meelis +11 · 164 citations
Computer Science · Mathematics · #Algorithm #Artificial intelligence #Artificial neural network #Binary number #Brier score #Calibration #Computer science #Dirichlet distribution #Explainable Artificial Intelligence (XAI) #Extrapolation #Machine Learning and Data Classification #Mathematics #Multiplicative function #Neural Networks and Applications #Pattern recognition (psychology) #Probabilistic logic #Range (aeronautics) #Scaling #Softmax function #Statistics #cs.LG #stat.ML

paper · pdf · doi:10.48550/arxiv.1910.12656

published in arXiv (Cornell University) (Cornell University) · Accepted for presentation at NeurIPS 2019

openalex created_date 2019/09/05 · arxiv created 2019/10/28 · openalex publication_date 2019/10/28 · arxiv updated 2019/10/29 · openalex updated_date 2026/07/28

Abstract

Class probabilities predicted by most multiclass classifiers are uncalibrated, often tending towards over-confidence. With neural networks, calibration can be improved by temperature scaling, a method to learn a single corrective multiplicative factor for inputs to the last softmax layer. On non-neural models the existing methods apply binary calibration in a pairwise or one-vs-rest fashion. We propose a natively multiclass calibration method applicable to classifiers from any model class, derived from Dirichlet distributions and generalising the beta calibration method from binary classification. It is easily implemented with neural nets since it is equivalent to log-transforming the uncalibrated probabilities, followed by one linear layer and softmax. Experiments demonstrate improved probabilistic predictions according to multiple measures (confidence-ECE, classwise-ECE, log-loss, Brier score) across a wide range of datasets and classifiers. Parameters of the learned Dirichlet calibration map provide insights to the biases in the uncalibrated model.

Cited by

Related