Develops confidence intervals for ECE, a measure of model calibration.
problem Ensuring the calibration of probabilistic predictions in machine learning models.
method Develops confidence intervals for the ℓ2 Expected Calibration Error (ECE), considering top-1-to-k calibration. result Shows asymptotic normality and different convergence rates for calibrated and miscalibrated models, developing methods to construct valid confidence intervals.
Paper analyzes ECE bias and provides bounds for its estimation.
problem Understanding the estimation bias in ECE for machine learning models.
method Information-theoretic approach to analyze bias in uniform mass and uniform width binning strategies.
result Established upper bounds on ECE estimation bias and optimal number of bins.
New decision-theoretic calibration error metric improves prediction reliability.
problem Improving the reliability of predictions for decision-making.
method Proposed Calibration Decision Loss (CDL) and an efficient algorithm to achieve near-optimal CDL.
result Near-optimal CDL guarantees vanishing payoff loss from miscalibration.
Simple algorithm achieves distance to calibration error of at most 2√T+1.
problem Achieving distance to calibration error of O(√T) in adversarial setting.
method An extremely simple, efficient, deterministic algorithm.
result Obtains distance to calibration error at most 2√T+1.
Paper proposes a method to predict MOBA game winners with calibrated confidence.
problem Predicting MOBA game winners with noisy data and uncertain noise.
method A novel confidence-calibration method considering data uncertainty.
result Achieves outstanding expected calibration error (ECE) of 0.57%.
The paper addresses poor calibration in fine-tuned LLMs after preference alignment.
problem Poor calibration in fine-tuned Large Language Models (LLMs) after preference alignment.
method Proposes a calibration-aware fine-tuning approach to restore calibration without compromising model performance.
result Demonstrates the effectiveness of the proposed methods through extensive experiments.
This post introduces model calibration and evaluation measures, highlighting issues with a common measure.
problem Ensuring model confidence accurately reflects true outcomes.
method Explains common calibration definition, ECE, and its drawbacks.
result New evaluation measures needed for comprehensive model calibration.
A new framework evaluates LLM calibration in open-ended QA.
problem Evaluating LLM calibration in open-ended QA settings.
method Sem-ECE framework: sampling answers, grouping by semantic classes, and using frequencies as confidence.
result Sem-ECE estimators are unbiased and Sem2 achieves smaller calibration error. EC method calibrates neural networks by matching average confidence to correct label proportion.
problem Overoptimism in neural network prediction confidence.
method Expectation consistency (EC) post-training rescaling of weights.
result EC achieves similar calibration performance to temperature scaling (TS) but is based on a principled Bayesian principle.
Post-processing predictors reduces calibration errors for decision-making.
problem Predictors with low calibration error for machine learning may have high error for decision-making.
method Post-processing with ε distance to calibration adds noise to make predictions differentially private.
result Post-processing achieves O(√ε) ECE and CDL, asymptotically optimal.
This paper argues against using calibration metrics for assessing posterior probabilities and proposes expected proper scoring rules instead.
problem The assessment of posterior probabilities generated by machine learning classifiers using calibration metrics is flawed and should be replaced with expected proper scoring rules.
method The paper reviews proper scoring rules from a practical perspective, explains why expected PSRs are a principled measure of posterior quality, and introduces a new calibration metric called calibration loss.
result Calibration loss is superior to expected calibration error and expected score divergence calibration metrics for assessing posterior probabilities.
New method improves calibration of neural networks by targeting robust margins and local smoothness.
problem Poor calibration of neural networks, leading to unreliable confidence estimates.
method Intervene on training procedure by targeting robust margins and local smoothness.
result Improved out-of-sample calibration without sacrificing accuracy.
T-Cal tests model calibration with a minimax optimal test.
problem Detecting mis-calibration of predictive models using a finite validation dataset.
method T-Cal is a minimax optimal test for calibration based on a debiased plug-in estimator of the ℓ2-Expected Calibration Error (ECE). result T-Cal is a practical tool for testing the calibration of probabilistic classification methods.
A new calibration metric bridges testability and actionability.
problem Combining testability and actionable insights for forecast probabilities.
method Cutoff Calibration Error (CCE) that assesses calibration over intervals of forecasted probabilities.
result Cutoff Calibration Error is both testable and actionable.
Proposes CCE to assess point-wise reliability of neural network predictions.
problem Overconfidence and misaligned predictive distributions in neural networks.
method Introduces Conditional Congruence (CCE) metric using conditional kernel mean embeddings.
result CCE exhibits correctness, monotonicity, reliability, and robustness in high-dimensional regression tasks.
New method calibrates confidence for object detection and segmentation models.
problem Intrinsically miscalibrated confidence estimates in object detection and segmentation models.
method Introduces multivariate confidence calibration for object detection and segmentation, extending ECE.
result Improves calibration, positively impacts segmentation quality.
This paper rethinks confidence calibration under covariate shifts.
problem Calibration methods struggle with covariate shifts and unstable importance weighting.
method Derives Expectation consistency condition and proposes Expectation consistency loss (ECL).
result ECL loss is compatible with various types of calibration and has the same sample complexity as ECE.
Study shows sample complexity for multicalibration is Θ(ε^-3) with polylogarithmic factors.
problem Minimizing Expected Calibration Error (ECE) for predictors with respect to a family of groups.
method Proved necessary and sufficient sample complexity of Θ(ε^-3) for multicalibration, using online-to-batch reduction and lower bounds.
result Sample complexity of multicalibration is Θ(ε^-3) with polylogarithmic factors, distinguishing it from marginal calibration.
Leaky ReLU activations improve the calibration of Bayesian neural networks.
problem Bayesian neural networks struggle with mean-field variational inference for ReLU activations.
method Investigated the effect of activation functions on the calibration of Bayesian neural networks.
result Leaky ReLU activations lead to more Gaussian-like weight posteriors and lower expected calibration error.
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.
Bayesian Federated Learning improves model reliability in dynamic environments.
problem Uncertainty quantification and robust adaptation in distributed learning.
method Proposes a continual BFL framework using SGLD for sequential updates and continual learning challenges.
result Continual Bayesian updates preserve knowledge and adapt to evolving data.
New methods reduce bias in estimating calibration error.
problem Reducing bias in estimating calibration error.
method Synthesizing model outputs and using equal-mass bins.
result Two reliable calibration-error estimators found: debiased estimator and ECE_sweep.
New metrics improve uncertainty estimation on graph data.
problem Current GNNs focus only on nodewise scores, limiting uncertainty estimation.
method Proposed edgewise metrics for uncertainty estimation on graphs.
result GNN models with structured prediction perform better in uncertainty estimation.
A new perfectly truthful calibration measure improves prediction reliability.
problem Improving the reliability of predictions by ensuring they are conditionally unbiased.
method Designing a simple, perfectly truthful calibration measure called ATB.
result ATB is the first perfectly truthful calibration measure in the batch setting.
ABC improves uncertainty quantification in LLMs for clinical diagnostics.
problem Overconfident and poorly calibrated estimates of LLMs in clinical domains.
method Approximate Bayesian Computation (ABC) for likelihood-free inference.
result Improves accuracy by up to 46.9%, reduces Brier scores by 74.4%, and enhances calibration.
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.
Deep ensembles don't necessarily improve calibration in low data regimes.
problem Calibration issues in deep learning models, especially in low data regimes.
method Examination of data-augmentation, ensembling, and post-processing calibration methods.
result Standard ensembling techniques can lead to less calibrated models in low data regimes.
This work evaluates uncertainty in deep Gaussian processes.
problem Uncertainty quantification in deep Gaussian processes.
method Hierarchical deep Gaussian processes (DGPs) and Deep Sigma Point Processes (DSPPs) evaluated on regression and classification tasks.
result DSPPs provide strong in-distribution calibration but are less robust under distribution shift compared to ensembles.
DCC separates marginal estimation from dependence modeling for improved classification accuracy.
problem Classifying with strong dependence between features.
method Deep Copula Classifier using neural copula densities.
result Achieves excess-risk O(n−r/(2r+d)) for r-smooth copulas. Study on deep neural networks using concentration inequalities and optimal stopping.
problem Understanding the performance and structure of stochastic deep neural networks.
method Introduced concentration inequalities for SDNN outputs and an EC classifier. Determined the optimal number of layers via an optimal stopping procedure.
result Optimal number of layers for SDNNs determined via an optimal stopping procedure.
Approximations of loopy belief propagation, including expectation propagation and approximate message passing, have attracted considerable attention for probabilistic inference problems. This paper proposes and analyzes a generalization of Opper and Winther's expectation consistent (EC) approximate inference method. Th…
Applications such as weather forecasting and personalized medicine demand models that output calibrated probability estimates---those representative of the true likelihood of a prediction. Most models are not calibrated out of the box but are recalibrated by post-processing model outputs. We find in this work that popu…
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…
New truthful calibration errors improve model ranking in multiclass prediction.
problem Non-truthful calibration errors can mislead model comparisons.
method Introduced perfectly truthful calibration errors for multiclass predictions.
result Truthful calibration errors preserve decision-theoretic dominance and stabilize model rankings.
We propose a new framework to improve the calibration of neural networks.
problem Improving the accuracy of model confidence predictions.
method Introducing a differentiable surrogate for expected calibration error (DECE) and a meta-learning framework to optimise model hyper-parameters for validation set calibration.
result Achieved competitive performance with existing calibration approaches.
Improved image learning using elliptically contoured tensor-variate distributions.
problem Inadequate statistical analysis for tensor-valued data, especially with heavier or lighter tails.
method Developed a family of elliptically contoured tensor-variate distributions and derived their properties and procedures for estimation.
result Tensor-variate classification rules and tensor-on-tensor regression better predict and characterize data than TVN-based methods.
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.
Corrected misstatements about invariant rank in ECS manifold papers.
problem Misstated invariant rank (d=1 instead of d=2) in several papers on ECS manifolds.
method Identified the class of examples with d=1, explained why d=2, and listed corrections.
result Corrected papers to accurately state d=2 for ECS manifolds.
BYOV combines SSL and Bayesian methods for uncertainty estimation.
problem Model uncertainty in applications.
method Combines Bootstrap Your Own Latent (BYOL) and Bayes by Backprop (BBB).
result BYOV improves model calibration and reliability with various augmentations.
ECS evaluates synthetic CXR images' distributional fidelity.
problem Evaluating synthetic CXR images' distributional fidelity under privacy constraints.
method Characteristic function transforms of feature embeddings.
result ECS uncovers clinically relevant distributional discrepancies.
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.
A new algorithm reduces regret in multi-player bandits without collision info.
problem Decentralized multi-player multi-armed bandits with no collision info.
method EC-SIC algorithm using optimal error correction coding for reward statistics.
result Regret approaches that of centralized with collision info.
BoC probe assesses neural network confidence coherence, revealing architecture-specific uncertainty.
problem Poor calibration and OOD detection in neural networks.
method Bag-of-Coins (BoC) probe compares softmax confidence to pairwise dominance probabilities.
result BoC reveals clear ID/OOD separation for some architectures but not others.
EFDA extends LDA to non-Gaussian models using exponential families.
problem Classifying non-Gaussian data with LDA's limitations.
method EFDA uses exponential families to derive closed-form estimators for natural parameters and a linear decision rule.
result EFDA matches LDA's accuracy while reducing ECE by 2-6x, proving asymptotic calibration and efficiency.
We extend a variational framework to estimate calibration errors for Lp divergences.
problem Ensuring predicted probabilities match observed class frequencies in machine learning.
method Extend variational framework to Lp divergences, separating over- and under-confidence. result Avoids overestimation and separates over- and under-confidence.
New metrics CWSA and CWSA+ improve model evaluation under confidence thresholds.
problem Lack of metrics capturing model reliability under confidence thresholds.
method Introducing CWSA and CWSA+ metrics that reward confident accuracy and penalize overconfident mistakes.
result CWSA and CWSA+ outperform classical metrics in trust-sensitive tests.
New calibration measure SSCE ensures truthful prediction, unlike existing measures.
problem Ensuring truthful calibration measures in sequential prediction.
method Introduced a new calibration measure, Subsampled Smooth Calibration Error (SSCE).
result SSCE ensures truthful prediction, while existing measures are far from truthful.
In safety-critical applications a probabilistic model is usually required to be calibrated, i.e., to capture the uncertainty of its predictions accurately. In multi-class classification, calibration of the most confident predictions only is often not sufficient. We propose and study calibration measures for multi-class…