PaperlayerView exact revision on arXiv

Source-linked AI summary

Calibration in Deep Learning: A Survey of the State-of-the-Art

Cheng Wang

arXiv:2308.01222v4cs.LGcs.AI

TL;DR

Highly capable neural networks can still produce poorly calibrated confidence estimates, although real-world decisions often rely on confidence as well as correctness. This survey reviews calibration methods and measurement, including work on LLM calibration, and concludes that challenges remain, particularly for sequence generation.

  • Problem

    Highly accurate neural networks can be poorly calibrated, while real-world decisions often depend on confidence as well as correctness.

    Source: §1 · p. 1

  • Method

    The survey reviews calibration methods, their principles and trade-offs, and recent approaches for large language models.

    Source: §4.5.1 · p. 18, §5.2 · pp. 22–23, §7 · p. 24

  • Results

    The survey identifies continuing challenges and an under-explored need to extend calibration work to a broader range of generative models.

    Source: §7 · p. 24, §7.2 · p. 24

  • Takeaways & Limitations

    Because real-world decisions use confidence as well as correctness, calibrated confidence is relevant to safety-critical applications.

    Source: §1 · p. 1

    Binning-based ECE can bias calibration assessment by losing fine-grained information about data relationships.

    Source: §2.2.4 · p. 6

Abstract

from arXiv · show

Calibrating deep neural models plays an important role in building reliable, robust AI systems in safety-critical applications. Recent work has shown that modern neural networks that possess high predictive capability are poorly calibrated and produce unreliable model predictions. Though deep learning models achieve remarkable performance on various benchmarks, the study of model calibration and reliability is relatively under-explored. Ideal deep models should have not only high predictive performance but also be well calibrated. There have been some recent advances in calibrating deep models. In this survey, we review the state-of-the-art calibration methods and their principles for performing model calibration. First, we start with the definition of model calibration and explain the root causes of model miscalibration. Then we introduce the key metrics that can measure this aspect. It is followed by a summary of calibration methods that we roughly classify into four categories: post-hoc calibration, regularization methods, uncertainty estimation, and composition methods. We also cover recent advancements in calibrating large models, particularly large language models (LLMs). Finally, we discuss some open issues, challenges, and potential directions.

1. Introduction

Deep models can make accurate predictions while their confidence remains poorly calibrated, even though real-world decisions often depend on that confidence. This survey reviews calibration methods and principles, with particular attention to recent approaches and large language models.

1. Introduction

  • High predictive accuracy does not ensure calibrated probabilities: predicted class probabilities may fail to reflect the true likelihood of correctness.Calibration addresses both overconfident and underconfident predictions.Source: §1 · pp. 1–2
  • Real-world decisions often depend on prediction confidence, so uncalibrated confidence can compromise uncertainty assessments and lead to flawed decisions.This matters especially in safety-critical applications such as medical diagnosis, autonomous driving, and finance.Source: §1 · pp. 1–2

1.1 Scope and Focus

  • The survey addresses a gap in comprehensive reviews by emphasizing recent calibration methods, including approaches rarely covered in earlier surveys.Examples include kernel-based methods, differentiable calibration proxies, and meta-learning approaches.Source: §1.1 · p. 3
  • It explains method principles through relationships among over-parameterization, over-fitting, and over-confidence, and organizes methods into four categories.The categories are post-hoc, regularization, uncertainty estimation, and hybrid methods.Source: §1.1 · p. 3
  • The survey also examines calibration of large pre-trained models, particularly LLMs, and identifies future directions for the field.It notes growing interest in calibrating LLMs for zero-shot inference.Source: §1.1 · p. 3

2. Preliminaries and Backgrounds

Calibration concerns whether predicted probabilities correspond to the true likelihood of correctness; miscalibration is associated with model capacity, data conditions, training choices, and evaluation bias.

2.1 Definitions

  • Calibration measures how well predicted probabilities match the true likelihood of correctness; argmax calibration considers only the maximum predicted probability.Perfect calibration is difficult to obtain, and deviations from it constitute miscalibration.Source: §2.1 · p. 4

2.2.1 MODEL COMPLEXITY

  • High model capacity can increase calibration errors through overfitting, with overconfidence emerging as test negative log-likelihood rises.This pattern becomes more severe when high-capacity models are trained with limited data.Source: §2.2.1 · p. 4, §2.2 · p. 4

2.2.2 DATA ISSUES

  • Scarce or problematic data can undermine calibration: noise, imbalance, and dataset bias are linked to biased predictions or miscalibration.Data augmentation and pretrain-finetune training are described as ways to alleviate overfitting from scarce data; focal loss is reported effective for imbalanced datasets.Source: §2.2.2 · pp. 4–6

2.2.3 LEARNING OBJECTIVE AND REGULARIZATION

  • Training objectives and regularization affect calibration: optimizing accuracy alone can yield poorly calibrated probabilities, and models may be less calibrated without L2 regularization.Focal loss is described as promising for calibration, while several newer regularization techniques are cited as proposed approaches.Source: §2.2.3 · p. 6

2.2.4 BIAS IN EVALUATION METRICS

  • Binning-based evaluation metrics such as ECE can bias calibration assessment by losing fine-grained information about the data distribution.Adaptive ECE and Smooth ECE are noted as techniques proposed to address bias from binning.Source: §2.2.4 · p. 6

3. Calibration Measurements

Because exact calibration measurement is impossible with finite samples and continuous confidence values, the section surveys scalar metrics and reliability diagrams that approximate or visualize calibration.

3.1 Expected Calibration Error (ECE)

  • ECE summarizes the weighted average absolute gap between predicted probabilities and observed frequencies across bins.The section frames bin-based metrics as approximations because confidence is continuous and exact measurement with finite samples is impossible.Source: §3.1 · p. 7, §3 · p. 7

3.2 Maximum Calibration Error (MCE)

  • MCE reports the largest within-bin discrepancy, highlighting worst-bin error but offering a less comprehensive overall assessment than ECE.The section notes its particular relevance in high-risk applications requiring reliable confidence measures.Source: §3.2 · p. 7

3.3 Brier Score

  • The Brier Score measures both discrimination and calibration using mean squared probability error, and can be decomposed into reliability, resolution, and uncertainty.For multiclass problems, the score extends across classes.Source: §3.3 · pp. 7–8

3.4 Classwise ECE (CECE)

  • Classwise ECE computes calibration error separately for each class’s bins and averages across classes.It is presented as a macro-averaged extension of bin-based ECE.Source: §3.4 · p. 8

3.5 Adaptive ECE (AECE)

  • Adaptive ECE uses adaptive bin intervals so that each bin contains an equal number of samples, addressing bias from predefined bin sizes.The method sets intervals using sorted predictions.Source: §3.5 · p. 8

3.6 Reliability Diagram

  • A reliability diagram plots binned predicted probabilities against observed outcome frequencies to show how confidence aligns with empirical outcomes.The diagonal marks perfect calibration; points above it indicate overconfidence and points below it indicate underconfidence.Source: §3.6 · p. 8

4. Calibration Methods

The section groups calibration methods by approach and notes that hybrid methods combine techniques.

4. Calibration Methods

  • Calibration methods are grouped into post-hoc, regularization, uncertainty estimation, and hybrid categories.Source: §4 · p. 9
  • Hybrid methods combine different calibration methods.Source: §4 · p. 9
  • Figure 3 summarizes and categorizes the calibration methods discussed in the section.Source: §4 · p. 9

4.1 Post-hoc Methods

Post-hoc methods calibrate models after training, often by learning a map from validation data. The section reviews temperature scaling and extensions that adjust calibration for different methods and settings, alongside practical limits.

4.1.1 TEMPERATURE SCALING

  • Temperature scaling uses one positive parameter to adjust softmax sharpness, with its optimal value selected by validation negative log-likelihood.It simplifies class-wise scaling and offers good calibration with minimum computational complexity (Guo et al., 2017; Minderer et al., 2021a).Source: §4.1.1 · p. 9

4.1 Post-hoc Methods

  • Post-hoc methods learn a calibration map on validation data after training, using approaches such as histogram binning, isotonic regression, and Platt scaling.Temperature scaling and its extensions are among the parametric approaches.Source: §4.1.2 · p. 9, §4.1 · p. 9

4.1.2 TEMPERATURE SCALING EXTENSIONS

  • Dirichlet calibration assumes probability distributions are parameterized by Dirichlet distributions and offers a unique, interpretable map family coinciding with Beta calibration (Kull, Filho, & Flach, 2017).Source: §4.1.2 · pp. 9, 11
  • Attended TS selects class-aware validation samples (Mozafari et al., 2018); BTS uses equal-size confidence bins, while LTS learns local pixel- or voxel-specific temperatures for segmentation (Ding et al., 2021).Source: §4.1.2 · p. 11
  • For learning to defer, one-vs-all sigmoid outputs and a consistent surrogate yield calibrated expert-correctness probabilities while maintaining consistency with the 0–1 objective (Verma and Nalisnick, 2022).Across evaluated tasks, the approach attains calibration without sacrificing—and often improving—accuracy over the softmax baseline.Source: §4.1.2 · p. 11
  • Top-versus-All converts multiclass confidence calibration into a binary correctness task, improving ECE while preserving predicted labels and plug-in ease in text-based models (Le Coz et al., 2024).Source: §4.1.2 · pp. 11–12
  • Density-Aware Calibration uses feature-space k-nearest-neighbor density to rescale logits before a standard calibrator, reducing ECE under corruptions and OOD while preserving in-domain accuracy (Tomani et al., 2023).Source: §4.1.2 · p. 12

4.1.3 PRACTICAL CONSIDERATIONS

  • Post-hoc methods depend on representative validation data, and limited map capacity may reduce effectiveness under domain shift or nuanced prediction tasks.More flexible extensions can add complexity and require domain-specific tuning.Source: §4.1.3 · p. 12

4.2 Regularization Method

Regularization methods incorporate calibration objectives into training, either through loss penalties, differentiable calibration proxies, or data augmentation. Their calibration benefits vary with method and can involve tuning or computational trade-offs.

4.2.1 EXPLICIT REGULARIZATION

  • Explicit penalties such as L2 regularization can improve calibration, while entropy regularization discourages overly confident, peaked predictions (Guo et al., 2017; Pereyra et al., 2017).Liu et al. (2022) also interpreted label smoothing, focal loss, and confidence penalties as approximations of a linear penalty and proposed adding a margin.Source: §4.2.1 · pp. 12–13

4.2.2 IMPLICIT REGULARIZATION: FOCAL LOSS AND ITS EXTENSIONS

  • Focal loss can trade off KL-divergence minimization and entropy maximization; FLSD schedules γ by confidence range and improved calibration on in-domain and OOD data.Finding one suitable γ for all samples is challenging.Source: §4.2.2 · p. 13

4.2.3 IMPLICIT REGULARIZATION: DIFFERENTIABLE CALIBRATION PROXY

  • Differentiable proxies approximate calibration errors for training; kernel-based MMCE uses a differentiable calibration measure, while DECE guides a meta-network’s weight-decay coefficients.MDCA is a bin-free differentiable loss inspired by SCE, and its combination with focal loss had the lowest calibration error among evaluated methods.Source: §4.2.3 · pp. 13–14
  • Combining MDCA with focal loss produced the lowest calibration error among the evaluated methods.MDCA is differentiable and does not rely on binning.Source: §4.2.3 · p. 14

4.2.4 IMPLICIT REGULARIZATION: DATA AUGMENTATION

  • Label smoothing softens hard labels, while mixup combines random input-label pairs; mixup-trained models were better calibrated and less overconfident on OOD and noisy data.Mixing features alone did not bring calibration benefits; adding label smoothing could improve calibration.Source: §4.2.4 · p. 15

4.2.5 PRACTICAL CONSIDERATIONS

  • Regularization methods calibrate during training, but kernel evaluations and meta-network optimization can increase training time, and proxy losses may require tuning to avoid overfitting.Data augmentation adds hyperparameter and training-time costs and can occasionally underfit when over-applied.Source: §4.2.5 · pp. 15–16

4.3 Uncertainty Estimation

Uncertainty-estimation approaches introduce stochasticity or adjust gradients to improve calibration. They can improve reliability, but methods requiring multiple passes or model instances incur greater compute and memory costs.

4.3 Uncertainty Estimation

  • Uncertainty-estimation methods include Bayesian neural networks, ensembles, MC dropout, and Gumbel-softmax approaches that inject randomness in different ways.These approaches often use multiple forward passes to approximate predictive uncertainty.Source: §4.3 · p. 16, §4.3.2 · p. 16

4.3.1 BAYESIAN NEURAL NETWORKS

  • Bayesian neural networks approximate the intractable posterior over weights by minimizing KL distance, and estimate predictive uncertainty by averaging predictions across T inference runs.The posterior approximation can also be treated as maximizing the ELBO.Source: §4.3.1 · p. 16

4.3.2 STOCHASTIC AND GRADIENT-BASED UNCERTAINTY METHODS

  • BSCE-GRA scales gradients using Brier Scores across logits to prioritize uncertain samples, with reported competitive calibration while preserving model accuracy.The method adjusts gradient magnitudes without hindering convergence.Source: §4.3.2 · p. 17

4.3.3 PRACTICAL CONSIDERATIONS

  • Multiple forward passes or model instances increase compute and memory costs, limiting uncertainty-estimation methods in real-time or resource-constrained environments.The survey characterizes these methods as useful for calibration and uncertainty quantification when efficiency is not the primary constraint.Source: §4.3.3 · p. 17

4.4 Hybrid Calibration Methods

Hybrid calibration combines complementary methods, from temperature scaling paired with training-time approaches to uncertainty-aware conformal pipelines. The survey describes ETS, temperature scaling with MC dropout, and trade-offs in complexity and inference cost.

4.4 Hybrid Calibration Methods

  • Calibration compositions can pair post-hoc methods with regularization or uncertainty estimation; label smoothing combined with mixup training significantly improved calibration.Temperature scaling after regularization or implicit calibration is one example of this composition strategy.Source: §4.4 · p. 17

4.4.1 ENSEMBLE TEMPERATURE SCALING (ETS)

  • ETS adds expressivity to temperature scaling while retaining its accuracy-preserving and data-efficient properties.Its ensemble combines temperature-scaled, uncalibrated, and uniform predictions.Source: §4.4.1 · p. 17

4.4.2 TEMPERATURE SCALING WITH MC DROPOUT

  • Temperature scaling with MC dropout inserts a learnable positive temperature before softmax and optimizes it against validation-set NLL over multiple forward passes.The entropy of the resulting softmax likelihood represents uncertainty across classes.Source: §4.4.2 · pp. 17–18

4.4.3 CONFORMAL PREDICTION WITH MC DROPOUT

  • The MC-Dropout conformal pipeline achieves near-nominal coverage, acceptable interval widths, and improved interval scores across multiple datasets while supporting online calibration.It uses dropout samples to compute nonconformity scores and conformal quantiles to construct confidence intervals.Source: §4.4.3 · p. 18

4.4.4 PRACTICAL CONSIDERATIONS

  • Hybrid methods may increase complexity, interpretability challenges, and computational overhead, including extra tuning for ETS and multiple stochastic passes for MC Dropout.These costs can affect inference time and deployment suitability.Source: §4.4.4 · p. 18

4.5 Discussion and Analysis

The discussion compares calibration categories by their strengths and limitations, then surveys commonly used backbones and datasets. Method-to-method variation in these choices complicates comparisons with prior work.

4.5 Discussion and Analysis

  • The discussion analyzes representative methods’ strengths and limitations and outlines benchmark datasets and backbone models used in the literature.Table 1 summarizes strengths and limitations across calibration categories.Source: §4.5 · p. 18, Table 1 · p. 19

4.5.1 SUMMARY OF STRENGTHS AND LIMITATIONS

  • Beyond post-hoc calibration, methods generally incur greater architectural and computational complexity, with a rough ranking from hybrid to uncertainty estimation, regularization, and post-hoc methods.The exact comparison depends on the specific method selected.Source: §4.5.1 · p. 18

4.5.2 BACKBONE MODELS AND DATASETS

  • Calibration studies use varying datasets and backbones, making comparisons with prior methods challenging; the survey lists representative choices to support future baseline selection.Common image-classification backbones include ResNet, WideResNet, and DenseNet; LSTM and GP-CNN are widely used in NLP tasks.Source: §4.5.2 · pp. 19–20

5. Calibrating Pre-trained Large Models

The survey reviews calibration of large vision and language models, including evidence on distribution shift and prompt-based inference. It covers methods for correcting bias and calibrating confidence at both response and claim levels.

5.1 Large Vision Models (LVMs)

  • For vision models, better in-distribution calibration was associated with better OOD calibration, while model size and pretraining amount did not fully explain earlier calibration decay under distribution shift or model size.The section also reports that CLIP is miscalibrated in zero-shot settings and that an auxiliary-task temperature can be applied across prompts or datasets.Source: §5.1 · p. 20

5.2 Large Language Models (LLMs)

  • Prompt-based LLM calibration addresses confidence misalignment, bias, and black-box settings where internal logits are unavailable, especially in few-shot and zero-shot use.These constraints motivate work on reducing overconfidence and improving reliability across tasks and domains.Source: §5.2 · p. 21
  • Contextual calibration estimates prompt-induced bias from a context-free input, while Domain-Context Calibration uses random in-domain text to estimate label-wise priors.PriDe is a label-free, inference-time method for token-bias debiasing with reported efficiency and cross-domain generalization.Source: §5.2 · pp. 21–22
  • Batch Calibration estimates contextual bias from real task inputs across a batch, then debiases individual predictions; earlier contextual methods can struggle with multi-sentence inputs, and PROCA may overfit.Other surveyed approaches include mixup training on synthetic examples and PROCA, which fits a Gaussian mixture model over outputs.Source: §5.2 · p. 22
  • A batch-based calibration method is reported as robust to prompt variations and as unifying Contextual Calibration and Domain Calibration under a framework with state-of-the-art calibration and accuracy across tasks using PaLM-2 and CLIP.The passage specifies that B denotes batch size.Source: §5.2 · p. 22
  • Thermometer consistently yields lower calibration error than temperature scaling or MC-based methods, preserves accuracy, and runs faster through single-pass inference.It uses a lightweight, task-agnostic auxiliary calibration network without changing LLM outputs or requiring extra inference at deployment.Source: §5.2 · pp. 22–23
  • Atomic calibration separates long-form answers into factual claims and estimates confidence for each; experiments found more accurate and interpretable estimates, with generative and discriminative signals together performing best.The framework introduces UCCE and QCCE to assess calibration at claim and response levels.Source: §5.2 · p. 23

6. Calibration Applications

The survey describes calibration applications in medical imaging, financial decision-making, and autonomous driving. Across these settings, it discusses uncertainty estimates and calibration as inputs to practical decisions.

6.1 Medical Imaging and Analysis

  • In medical imaging, calibrated predictions and uncertainty estimates support clinical decisions, while studies also report calibration and discrimination biases in skin-lesion detection.The survey describes label-noise robustness work based on expressing noisy-label confusion matrices through clean-label and noise confusion matrices.Source: §6.1 · p. 23

6.2 Financial Decision Making

  • In finance, calibrated risk estimates are used in contexts including fraud detection and financial forecasting to inform risk management and investment decisions.The survey notes applications to transaction-risk estimates and portfolio allocation.Source: §6.2 · pp. 23–24

6.3 Autonomous Driving

  • For autonomous vehicles, MC-dropout has been applied to quantify object-detection uncertainty, with calibration performed based on that uncertainty analysis.The survey frames calibrated predictions as relevant to safe navigation in complex traffic scenarios.Source: §6.3 · p. 24

7. Conclusion and Future Work

The survey concludes that calibration has advanced but still faces measurement challenges and under-explored aspects. It highlights calibration bias and sequence-generation calibration, particularly token-level probabilities, as directions for further work.

7. Conclusion and Future Work

  • The survey identifies remaining challenges and under-explored aspects in deep-model calibration despite recent advances.Source: §7 · p. 24

7.2 Calibrating Generative Models

  • Sequence-generation calibration is under-explored, and token-level miscalibration has been linked to counter-intuitive BLEU drops as beam size increases.Recalibrating token probabilities can improve sequence-level calibration.Source: §7.2 · pp. 24–25

7.1 Mitigating Calibration Bias

  • Increasing bin counts raises ECE and MCE for a given model, while calibration error remains sensitive to the binning scheme.KDE-based ECE estimation and Bias-by-Construction address data efficiency and bias modeling; future work includes unbiased estimation and bias–variance trade-offs.Source: §7.1 · p. 24, Figure 4 · p. 25

7.2 Calibrating Generative Models

  • Token-level miscalibration persists in LLM few-shot settings; the softmax bottleneck is especially problematic with large vocabularies, and errors can compound across generated tokens.Source: §7.2 · p. 25
  • Generative calibration, sequence likelihood calibration, and SLIC are promising starting points for developing calibration methods suited to sequence generation.Source: §7.2 · p. 25

Related papers

Loading 2308.01222v4…