The paper extends calibration to sets of probabilistic classifiers, finding many ensembles are poorly calibrated.
problem Evaluating the validity of epistemic uncertainty in sets of probabilistic classifiers.
method Proposed a novel nonparametric calibration test for sets of probabilistic classifiers.
result Ensembles of deep neural networks are often not well calibrated.
Classifier calibration does not always go hand in hand with the classifier's ability to separate the classes. There are applications where good classifier calibration, i.e. the ability to produce accurate probability estimates, is more important than class separation. When the amount of data for training is limited, th…
CCAC calibrates DNN classifiers on OOD datasets by separating mis-classified samples.
problem Calibrating DNN classifiers on out-of-distribution datasets is challenging.
method CCAC introduces an auxiliary class to map DNN output to calibrated confidence, separating mis-classified from correctly classified samples.
result CCAC consistently outperforms prior methods on various DNN models, datasets, and applications.
New framework for evaluating multiclass classifier calibration.
problem Ensuring classifiers are well-calibrated for trustworthy predictions.
method Utility Calibration framework that measures calibration error relative to a utility function.
result Unified and robust interpretation of existing calibration metrics.
Calibrating classifiers reduces grouping loss using sufficiency criteria.
problem Grouping loss in probabilistic classifier calibration is often overlooked.
method Revisited Langford & Zadrozny's probing reduction approach and introduced Brier curves.
result The probing reduction approach reduces grouping loss and supports sufficient calibration.
Accurate calibration of probabilistic predictive models learned is critical for many practical prediction and decision-making tasks. There are two main categories of methods for building calibrated classifiers. One approach is to develop methods for learning probabilistic models that are well-calibrated, ab initio. The…
Two methods improve calibration of probabilistic classifiers, especially for multi-class problems.
problem Improving calibration of probabilistic classifiers, especially for multi-class problems.
method Two techniques: reduced calibration and class-wise calibration.
result Class-wise reduced calibration algorithms reduce prediction and per-class calibration errors.
A new framework separates classifier calibration and discrimination.
problem Combining reliability and resolution in probabilistic predictions.
method Manokhin Probability Matrix separates reliability and resolution using Spiegelhalter Z-statistic and AUC-ROC.
result Classifiers are categorized into four archetypes: Eagle, Bull, Sloth, and Mole.
Meta-Cal improves post-hoc calibration of neural networks.
problem Improving the accuracy of uncalibrated neural network predictions.
method Meta-Cal uses a base calibrator and a ranking model with constraints to provide high-probability bounds.
result Meta-Cal significantly outperforms existing methods in post-hoc multi-class classification calibration.
Survey on assessing and improving classifier calibration for better decision making.
problem Ensuring classifiers correctly quantify prediction uncertainty.
method Overview of principles, methods, and evaluation metrics for calibration.
result New methods and extensions from binary to multiclass settings.
The paper addresses fairness issues in screening classifiers, proposing within-group monotonicity to avoid unfair treatment of qualified candidates.
problem Within-group unfairness in screening classifiers using calibrated models.
method Introducing within-group monotonicity as a property to avoid unfair treatment and developing an efficient post-processing algorithm based on dynamic programming.
result Within-group monotonicity can be achieved efficiently and often at a small cost, improving fairness without significantly compromising prediction accuracy.
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 exist…
Certified calibration methods protect model confidence from adversarial attacks.
problem Adversarial attacks degrade model calibration, reducing confidence in predictions.
method Developed certified calibration methods to provide worst-case bounds on calibration under adversarial perturbations.
result Certified calibration methods produce analytic and approximate bounds for the Brier score and expected calibration error.
This paper reviews metrics to assess AI model calibration accuracy.
problem AI model probabilities do not always match their true accuracy.
method Comprehensive review of 82 probability calibration metrics.
result Identified 4 classifier families and 1 object detection family of metrics.
Learning accurate probabilistic models from data is crucial in many practical tasks in data mining. In this paper we present a new non-parametric calibration method called \textit{ensemble of near isotonic regression} (ENIR). The method can be considered as an extension of BBQ, a recently proposed calibration method, a…
A framework assesses the trustworthiness of probabilistic classifiers using local calibration error.
problem Assessing the trustworthiness of probabilistic classifiers beyond traditional metrics.
method I-trustworthy framework linking local calibration to trustworthiness; Kernel Local Calibration Error (KLCE) method for hypothesis testing.
result The effectiveness of the proposed test statistic demonstrated through simulated and real-world datasets.
Proposes a new method for multi-class classification with well-calibrated predictions.
problem Improving the accuracy and reliability of multi-class classification models.
method Trains data in a latent space induced by an (n−1)-dimensional simplex, then extends and fits a regression model. result Demonstrates a well-calibrated classifier with improved prediction and calibration properties.
Study calibrates high-dimensional binary classifiers using angle between estimator and true weights.
problem Calibrating high-dimensional binary classifiers with provable properties.
method Interpolates with a chance classifier to construct well-calibrated predictor based on angle between estimator and true weights.
result Angular calibration approach is provably well-calibrated in high dimensions, minimizing Bregman divergence.
Many applications of classification methods not only require high accuracy but also reliable estimation of predictive uncertainty. However, while many current classification frameworks, in particular deep neural networks, achieve high accuracy, they tend to incorrectly estimate uncertainty. In this paper, we propose a …
A new method for measuring prediction uncertainty in classifiers.
problem Measuring uncertainty of predictions from machine learning methods.
method Density Based Calibration (DBCal) technique.
result Expected calibration error of less than 0.2% on binary classifiers and less than 3% on semantic segmentation networks.
Binary classification is highly used in credit scoring in the estimation of probability of default. The validation of such predictive models is based both on rank ability, and also on calibration (i.e. how accurately the probabilities output by the model map to the observed probabilities). In this study we cover the cu…
New method calibrates classifier probabilities with guaranteed coverage.
problem Inaccurate probability estimates by classifiers in high-risk applications.
method Adaptive temperature scaling algorithm for conformal prediction.
result Improves calibration error measures and standard metrics across various tasks.
Paper develops a new algorithm to improve screening processes.
problem Improving the efficiency and accuracy of screening processes.
method Developed a distribution-free screening algorithm called Calibrated Subset Selection (CSS).
result CSS provides near-optimal shortlists with provable diversity guarantees.
Optimizes calibration error estimators for better classifier trustworthiness.
problem Lack of guidance on selecting and tuning calibration error estimators.
method Reformulates calibration estimation as a regression problem with i.i.d. input pairs.
result Demonstrates the effectiveness of optimized calibration estimators on image classification tasks.
JUCAL jointly calibrates aleatoric and epistemic uncertainties in classifier ensembles.
problem Misrepresentation of predictive uncertainty due to unbalanced aleatoric and epistemic uncertainties.
method Joint Uncertainty Calibration (JUCAL) that jointly calibrates two constants to weight and scale uncertainties.
result Significantly outperforms state-of-the-art calibration methods across various text classification tasks.
Paper proposes CCVR to improve classifier calibration in federated learning with non-IID data.
problem Improving classifier calibration in federated learning with non-IID data.
method Proposes CCVR, a simple algorithm that adjusts the classifier using virtual representations.
result CCVR achieves state-of-the-art performance on popular federated learning benchmarks.
Probabilistic classifiers output a probability distribution on target classes rather than just a class prediction. Besides providing a clear separation of prediction and decision making, the main advantage of probabilistic models is their ability to represent uncertainty about predictions. In safety-critical applicatio…
Unified Uncertainty Calibration improves AI predictions by combining different types of uncertainty.
problem AI classifiers struggle with uncertainty, leading to miscalibrated predictions and poor performance.
method Unified Uncertainty Calibration (U2C) combines aleatoric and epistemic uncertainties to improve prediction quality.
result U2C outperforms traditional reject-or-classify methods across various ImageNet benchmarks.
Proposes and evaluates three diagnostic graphics for probabilistic classifiers.
problem Evaluating and comparing probabilistic classifiers.
method Triptych of diagnostic graphics: reliability diagram, ROC curve, Murphy diagram.
result Visual diagnostics reveal distinct aspects of forecast performance.
A new L2D system produces calibrated probabilities of expert correctness without sacrificing accuracy.
problem Calibration of learning to defer systems for safety.
method One-vs-all classifiers with a consistent surrogate loss function.
result Proposes a calibrated L2D system that outperforms existing methods in accuracy and calibration.
Study compares various calibration methods for binary classification tasks.
problem Improving probabilistic predictions in binary classification models.
method Benchmarked 21 classifiers using 5 calibration methods on real data.
result Venn-Abers predictors and Beta calibration show the largest log-loss reductions.
Proposes a new calibration error estimator for deep neural networks.
problem Improves calibration of deep neural networks, especially for canonical calibration.
method Uses a Dirichlet kernel density estimate to create a low-bias, trainable calibration error estimator.
result Asymptotically converges to true Lp calibration error, enabling efficient estimation and mini-batch updates. Mix-n-Match improves uncertainty calibration in deep learning.
problem Post-hoc calibration of machine learning classifiers.
method Ensemble and composition strategies to improve accuracy, efficiency, and expressive power.
result Mix-n-Match strategies achieve better data-efficiency and expressive power while maintaining classification accuracy.
This paper characterizes and designs loss functions for robust classification with abstention.
problem Ensuring robustness against adversarial attacks and knowing when to abstain from prediction.
method Proposes adversarial robust reject option loss and characterizes surrogates for calibration.
result Shifted Double Ramp Loss and Shifted Double Sigmoid Loss satisfy the calibration conditions.
TCE measures calibration error with a test-based approach.
problem Measuring calibration error of probabilistic binary classifiers.
method TCE uses a novel loss function based on a statistical test.
result TCE offers clear interpretation, consistent scale, and enhanced visual representation.
TS improves class coverage but reduces CP set size, offering a trade-off.
problem Combining temperature scaling with conformal prediction for deep classifiers.
method Empirical study and mathematical theory of TS's effect on CP.
result TS allows trading prediction set size and conditional coverage.
A popular methodology for building binary decision-making classifiers in the presence of imperfect information is to first construct a non-binary "scoring" classifier that is calibrated over all protected groups, and then to post-process this score to obtain a binary decision. We study the feasibility of achieving vari…
New method calibrates uncertainty estimates for image classifiers without labeled data.
problem Uncertainty estimates for modern classifiers are unreliable without labeled calibration data.
method Calibrates uncertainty estimates using unlabeled examples for distribution shifts.
result Proposes a method that provides excellent uncertainty estimates under natural distribution shifts.
The concept of refinement from probability elicitation is considered for proper scoring rules. Taking directions from the axioms of probability, refinement is further clarified using a Hilbert space interpretation and reformulated into the underlying data distribution setting where connections to maximal marginal diver…
Calibrated ensembles improve both ID and OOD accuracy in distribution shift.
problem Desired balance between in-distribution and out-of-distribution accuracy.
method Ensemble standard and robust models, calibrating on ID data only.
result ID-calibrated ensembles outperform state-of-the-art methods on multiple datasets.
The study quantifies decision-making risks from suboptimal classifiers and proposes methods to reduce these risks.
problem Excess risk in decision-making from suboptimal probabilistic classifiers.
method Analytical expressions and upper/lower bounds for excess risk, calibration curve estimation, grouping loss estimator.
result Identifies regimes where recalibration alone or post-training is more effective.
New method calibrates models under covariate shifts.
problem Calibration of models can be lost under covariate shifts.
method Importance sampling based approach.
result Efficacy demonstrated on real-world and synthetic datasets.
Obtaining accurate and well calibrated probability estimates from classifiers is useful in many applications, for example, when minimising the expected cost of classifications. Existing methods of calibrating probability estimates are applied globally, ignoring the potential for improvements by applying a more fine-gra…
Overconfidence and underconfidence in machine learning classifiers is measured by calibration: the degree to which the probabilities predicted for each class match the accuracy of the classifier on that prediction. How one measures calibration remains a challenge: expected calibration error, the most popular metric, ha…
The study examines when to trust confidence thresholding in pseudo-labelling regression.
problem Calibrated probabilities from classifiers used for pseudo-labelling need careful handling to avoid bias in downstream regression.
method Developed a diagnostic apparatus to predict and bound the bias induced by confidence thresholding, derived a closed-form expression for the attenuation bias.
result The bias can be predicted from the residual score variance V∗, motivating a structural separation between classifier features and downstream controls. We investigate the problem of multiclass classification with rejection, where a classifier can choose not to make a prediction to avoid critical misclassification. First, we consider an approach based on simultaneous training of a classifier and a rejector, which achieves the state-of-the-art performance in the binary …
New method estimates grouping loss in neural networks to improve confidence scores.
problem Improving confidence scores in neural networks to reflect true posterior probabilities.
method Proposed an estimator to approximate the grouping loss.
result Modern neural networks exhibit grouping loss, especially in distribution shifts.
Proposes top-label calibration and M2B framework for multiclass to binary calibration.
problem Multiclass calibration and interpretation issues.
method Top-label calibration and M2B reduction framework.
result M2B + HB achieves lower calibration error than other methods.