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.
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.
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.
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%.
Novel framework for unbiased confidence estimates in object detection.
problem Unbiased confidence estimates for safety-critical object detection.
method Combines regression output with additional information for calibration.
result Calibrated confidence estimates for image location and scale.
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.
This paper improves lottery ticketing by calibrating network confidence.
problem Uncalibrated confidence in lottery tickets leads to overconfidence and poor performance.
method The paper introduces various calibration strategies and explores their impact on lottery tickets.
result Calibration mechanisms consistently improve lottery ticket performance, even under distribution shifts.
Survey of methods to calibrate neural network predictions.
problem Ensuring neural networks provide accurate confidence levels.
method Empirical comparison of calibration methods.
result Various techniques for calibrating neural networks.
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.
Calibrates network confidence for unsupervised domain adaptation.
problem Calibrating a model trained on a source domain to a target domain without labeled data.
method Estimates network accuracy on the target domain and calibrates prediction confidence directly in the target domain.
result Significantly outperforms existing methods across standard datasets.
The paper tackles confidence calibration for exploratory machine learning problems.
problem Difficulty in curating datasets and confusion about category validity.
method Introduces four new algorithms for category-specific confidence estimation, including kernel density ratios.
result Kernel density ratios provide a novel approach to confidence calibration, especially for exploratory problems.
The paper investigates how dataset quality and heterogeneity affect model confidence in machine learning.
problem Understanding how dataset quality and heterogeneity impact model confidence in machine learning.
method The study uses theoretical explanations and experimental demonstrations to investigate the effects of dataset size, label noise, and class heterogeneity on model confidence.
result Label noise reduces model confidence, while reduced dataset size increases it, and class heterogeneity leads to inconsistent confidence across classes.
New method improves model calibration by adjusting confidence based on prediction correctness.
problem Improving model confidence alignment with true class probabilities.
method Post-hoc calibration objective using transformed samples for training.
result Competitive calibration performance on in-distribution and out-of-distribution test sets.
Confidence intervals and joint confidence sets are constructed for the nonparametric calibration of exponential Lévy models based on prices of European options. To this end, we show joint asymptotic normality in the spectral calibration method for the estimators of the volatility, the drift, the jump intensity and the …
LLMs show surprising confidence in their answers, beyond just tokens.
problem LLMs lack meaningful confidence estimates for their responses.
method Semantic calibration test based on local loss optimality and equivalence classes.
result Base LLMs are semantically calibrated across tasks, contrary to expectations.
Improves confidence calibration in neural networks by smoothing labels based on class similarity.
problem Improving confidence calibration in deep neural networks for safety-critical applications.
method Proposes a novel label smoothing technique where label values are based on similarities with the reference class, using different similarity measurements.
result Consistently outperforms state-of-the-art calibration techniques on various datasets and network architectures.
Study three types of uncertainty quantification for binary classification without distributional assumptions.
problem Uncertainty quantification for binary classification in a distribution-free setting.
method Established theorems connecting calibration, confidence intervals, and prediction sets for score-based classifiers.
result Distribution-free calibration is only possible using scoring functions that partition feature space into countably many sets.
Linguistic calibration improves long-form text confidence.
problem LMs hallucinate, leading to suboptimal decisions.
method Defining linguistic calibration, training framework, reinforcement learning.
result Llama 2 7B is significantly more calibrated than baselines.
Trained DNN models are increasingly adopted as integral parts of software systems, but they often perform deficiently in the field. A particularly damaging problem is that DNN models often give false predictions with high confidence, due to the unavoidable slight divergences between operation data and training data. To…
In classification applications, we often want probabilistic predictions to reflect confidence or uncertainty. Dropout, a commonly used training technique, has recently been linked to Bayesian inference, yielding an efficient way to quantify uncertainty in neural network models. However, as previously demonstrated, conf…
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.
Deep neural networks (DNNs) are poorly calibrated when trained in conventional ways. To improve confidence calibration of DNNs, we propose a novel training method, distance-based learning from errors (DBLE). DBLE bases its confidence estimation on distances in the representation space. In DBLE, we first adapt prototypi…
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. We propose an algorithm combining calibrated prediction and generalization bounds from learning theory to construct confidence sets for deep neural networks with PAC guarantees---i.e., the confidence set for a given input contains the true label with high probability. We demonstrate how our approach can be used to cons…
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. New methods improve confidence set calibration in complex models.
problem Challenges in maintaining confidence set coverage in complex models.
method TRUST and TRUST++ methods using simulated data for calibration.
result Methods achieve distribution-free conditional coverage and robust inference.
This paper improves confidence measurement in deep metric learning models.
problem Measuring confidence in deep metric learning models is challenging.
method Approximates class distributions using Gaussian kernel smoothing and calibrates the confidence metric.
result Improves generalization and robustness of deep metric learning models.
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.
We propose a generic framework to calibrate accuracy and confidence of a prediction in deep neural networks through stochastic inferences. We interpret stochastic regularization using a Bayesian model, and analyze the relation between predictive uncertainty of networks and variance of the prediction scores obtained by …
Post-hoc calibration improves uncertainty under domain shift.
problem Improving uncertainty calibration under domain shift.
method Apply perturbations to validation set before post-hoc calibration.
result Perturbation step results in better calibration under domain shift.
New method calibrates deep networks by preserving top-k predictions.
problem Calibrated confidence scores for multi-class deep networks to avoid rare mistakes.
method Intra order-preserving functions combined with neural network architecture.
result Outperforms state-of-the-art methods in evaluation metrics.
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.
The paper shows over-confidence in models isn't just due to over-parametrization.
problem Over-confidence in machine learning models, especially in binary classification.
method Theoretical analysis of logistic regression and other binary classification problems.
result Logistic regression is inherently over-confident in certain settings, but over-confidence is not always the case.
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.
Study on learning to defer to multiple experts with consistent surrogates and confidence calibration.
problem Addressing the open problems of consistent surrogates, confidence calibration, and ensembling of experts.
method Derive two consistent surrogates (softmax and OvA) and propose a conformal inference technique for choosing experts.
result The OvA-based loss does not cause mis-calibration propagation, while the softmax-based loss does.
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.
Annealing Double-Head calibrates deep neural networks during training.
problem Overestimation or underestimation of predictive confidence in deep neural networks.
method An additional calibration head and Annealing technique to dynamically scale logits.
result State-of-the-art model calibration performance achieved without post-processing.
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.
This paper improves multi-class calibration methods using mutual information maximization-based binning.
problem Calibration of deep neural network predictions, especially for small prior classes.
method I-Max concept for binning, shared class-wise calibration strategy.
result Improves multi-class ranking and calibration performance using a small calibration set.
CAT improves domain adaptation for fault diagnosis by calibrating teacher network predictions.
problem Performance drops in deep learning models when applied to different data distributions.
method CAT uses post-hoc calibration techniques to calibrate predictions of the teacher network during self-training.
result CAT achieves state-of-the-art performance on most transfer tasks in domain-adaptive IFD.
Time series foundation models are well-calibrated, improving over baseline models.
problem Calibration of time series foundation models for practical applications.
method Systematic evaluations of five time series foundation models and two baselines, assessing calibration, prediction heads, and long-term forecasting.
result Time series foundation models are consistently better calibrated than baseline models and do not show over- or under-confidence.
New method improves auto-labeling accuracy by optimizing confidence functions.
problem Overconfident model scores lead to poor TBAL performance.
method Developed a new post-hoc method, Colander, to optimize TBAL confidence functions.
result Achieves up to 60% improvement in coverage over baseline methods.
New framework assesses LM uncertainty without thresholding.
problem Uncertainty quantification for LMs, especially comparing different measures.
method Rank-Calibration framework to assess uncertainty and confidence measures.
result Higher uncertainty correlates with lower generation quality.
Spatially weighted conformal prediction improves uncertainty quantification in house price models.
problem Uncertainty quantification in automated valuation models with spatial dependencies.
method Survey and demonstration of various spatially weighted approaches to adjust conformal prediction confidence sets.
result Spatially weighted CP makes confidence sets more consistently calibrated across geographical regions.
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…
The study examines label smoothing to improve confidence calibration in fine-tuned LLMs.
problem Improving confidence calibration in fine-tuned large language models (LLMs) after instruction tuning.
method Examine various open-sourced LLMs, label smoothing, and custom kernel design.
result Label smoothing is effective in maintaining confidence calibration but faces challenges in large vocabulary LLMs.
We present a new method for computing ASR word confidences that effectively mitigates the effect of ASR errors for diverse downstream applications, improves the word error rate of the 1-best result, and allows better comparison of scores across different models. We propose 1) a new method for modeling word confidence u…
No policy can simultaneously be fully autonomous, optimally calibrated, and helpful, proving a trilemma.
problem Proving impossibility of a policy achieving maximum helpfulness, optimal calibration, and full autonomy.
method Geometric proof showing that adding any non-affine autonomy incentive to a strictly proper scoring rule destroys strict properness.
result The Behavioral Credibility Trilemma: no policy can achieve all three goals simultaneously.