Pith. sign in

REVIEW 4 major objections 5 minor 42 references

Conformal Uncertainty Indicator for Continual Test-Time Adaptation

T0 review · 4 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that a conformal uncertainty wrapper with a domain-shift-compensated threshold can restore coverage under continual domain shifts and reduce error accumulation in test-time adaptation.

desk verdict Plausible first application of conformal prediction to CTTA, but the coverage guarantee rests on an unreported tuned beta, so the central claim is not yet established. read the letter →

arxiv 2502.02998 v1 pith:X7WC2CDD submitted 2025-02-05 cs.LG

Fan Lyu , Hanyu Zhao , Ziqi Shi , Ye Liu , Fuyuan Hu , Zhang Zhang , Liang Wang This is my paper · ORCID
classification cs.LG
keywords continualtest-timeadaptationconformalpredictioncoveragegapdomainshiftuncertaintyestimationsetspseudo-labelweighting
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that conformal prediction, usually invalid when test data stop being exchangeable, can be repaired for continual test-time adaptation by compensating the conformal threshold with an online measured domain shift. It proposes a Conformal Uncertainty Indicator (CUI) whose prediction-set size marks each test prediction as reliable or not, then weights self-adaptation toward reliable predictions. During testing, CUI measures drift through a joint source-and-current-model distribution distance, so that error accumulation in the adapted model becomes visible in the uncertainty scores. On three corruption benchmarks, CUI brings coverage close to the user-specified level while lowering error rates of six existing CTTA methods. If true, CTTA models gain a model-agnostic uncertainty signal that helps them avoid trusting bad pseudo-labels, which matters for long-term autonomous driving and medical deployment.

What carries the argument

The load-bearing object is the compensated quantile threshold $\hat{\tau} = \tau^* - \beta\rho$. $\tau^*$ is the usual conformal quantile of calibration nonconformity scores $s = 1 - \pi(y_{\text{true}}|x)$, and $\rho$ is the sum over calibration/test pairs of Jensen-Shannon divergence between softmax vectors concatenated from the source and current models. The subtraction enlarges prediction sets as drift grows, restoring coverage; the prediction-set size is the uncertainty output, and a weight function $\gamma(x)$ converts that size into an adaptation weight for each test sample.

What would settle it

Fix $\beta$ and $\alpha$, run CUI on the same corruption set in several different orders (for example severity descending, or alternating easy and hard corruptions), and record per-domain coverage; if coverage systematically drops below $1-\alpha$ in any ordering, the linear compensation is not stable across domains. A single such ordering is enough to refute the method as stated.

Watch

Extended reading notes

Core claim

The central claim is that a model-agnostic conformal wrapper can provide trustworthy uncertainty for continual test-time adaptation if the threshold is lowered by $\beta\rho$, where $\rho$ summarizes domain and model drift and $\beta$ is a fixed compensation factor. CUI measures $\rho$ as the summed Jensen-Shannon divergence between joint representations of calibration and test batches under both the source and current models, so that error accumulation in the adapted model is visible in the score. Prediction sets are then formed by thresholding one-minus-softmax nonconformity scores at the compensated quantile, and set size serves as the uncertainty indicator. Weighting each sample's adaptation loss by its relative set size (larger weight for singleton sets, zero for empty sets) is claimed to reduce error accumulation and improve performance across Tent, CoTTA, SATA, RDumb, C-CoTTA, and RMT.

Load-bearing premise

The method's coverage repair rests on the assumption that a fixed scalar $\beta$ linearly maps the measured domain shift $\rho$ to the needed threshold correction, under every order and severity of domain changes; the paper gives no proof or bound for that relation.

Editorial extensions

If this is right

  • Existing CTTA methods can be wrapped with CUI and its adaptation strategy without architectural changes, since CUI only needs forward passes of the current model and a small labeled calibration set.
  • With user coverage levels $\alpha \in \{0.1, 0.2, 0.3\}$, measured coverage on the corruption benchmarks stays near the expected $1-\alpha$, whereas exchangeable and non-exchangeable baselines show visible coverage gaps.
  • Weighting the adaptation loss by prediction-set size lowers both error rate and inefficiency across Tent, CoTTA, SATA, RDumb, C-CoTTA, and RMT.
  • Privacy-first and efficiency-first ways of building the calibration set give similar performance, so the user can choose based on data availability.
  • Larger calibration storage improves conformal performance, and CUI with adaptation beats storing the same number of samples for source replay.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because $\beta$ is not fixed across datasets and the paper only shows that small $\beta$ under-compensates while large $\beta$ over-compensates, the practical guarantee rests on an offline tuning step; a testable extension is to estimate $\beta$ online from the running coverage error.
  • If the linear form $\hat{\tau} = \tau^* - \beta\rho$ is the right mechanism, per-domain coverage should track $\rho$ monotonically; checking monotonicity across the fifteen corruption types would separate a real effect from curve-fitting to a chosen $\beta$.
  • The same set-size signal could serve for abstention or human intervention in high-stakes deployment, not just for reweighting, since empty and large prediction sets already flag the most unreliable inputs.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes Conformal Uncertainty Indicator (CUI), a model-agnostic post-hoc module that applies conformal prediction to continual test-time adaptation. For each test input, CUI computes a nonconformity threshold from a labeled source calibration set and lowers it by beta times a joint JS-divergence score rho between calibration and test batches under source and current models, to compensate for coverage loss under domain shift. Prediction-set size is used as an uncertainty measure, and a weighting scheme (CPAda) up-weights reliable samples during adaptation. Experiments on CIFAR10-C, CIFAR100-C, and ImageNet-C combine CUI with Tent, CoTTA, SATA, RDumb, C-CoTTA, and RMT, reporting that CUI yields near-target coverage and that CPAda reduces error rates.

Significance. If the coverage-compensation mechanism were validated, CUI would be a practically useful wrapper: it is model-agnostic, requires only a small calibration set, and the CPAda weighting improves several existing CTTA methods, including strong recent baselines. The paper makes falsifiable cross-method empirical predictions and separately evaluates uncertainty quality (COV and INE) from adaptation accuracy, which is a strength. However, the central coverage claim currently rests on an unstated, possibly tuned beta and on an unnormalized shift score; until the authors report a fixed beta selection rule, normalize rho, and show that coverage holds with fixed hyperparameters across settings, the near-target COV values cannot be distinguished from fitted outcomes.

major comments (4)
  1. [Sec. 4.1.3, Eq. (8)] The central compensation step tau_hat = tau* - beta*rho is introduced without any derivation, bound, or monotonicity result linking rho to the coverage gap kappa in Eq. (3). Moreover, beta is never reported in Tables 1 or 2, and no selection rule is given; Fig. 3(b) shows only error-rate sensitivity on CIFAR100C and does not report the beta values used in the main tables. Because rho in Eq. (5) is an unnormalized sum over |C|*|B| pairwise JS divergences, its magnitude depends on calibration-set size (50/100/500 in the three benchmarks) and batch size, so a single beta cannot transfer across benchmarks; even within a benchmark, the reported near-target COV values can be explained by per-dataset tuning of beta rather than by the mechanism. The authors should report beta, normalize rho (e.g., divide by |C||B|), and validate with a fixed beta across all settings and alpha values.
  2. [Eq. (7)] The nonconformity score is written s(pi_{theta_crt}(x)) = 1 - y_hat, which is not well-defined: y_hat is an integer label, not a probability. It should be s(x,y) = 1 - pi_theta(y|x). Also, the calibration scores are computed with the current adapted model theta_crt, while the calibration labels come from the source distribution; the paper does not justify that these scores remain comparable to the source-model scores used to define the threshold, nor does it state when tau* is recomputed during the stream. This is load-bearing because both the threshold and the prediction-set construction depend on it.
  3. [Sec. 5.1, Tables 1-2] COV values are systematically 1-4 percentage points below the target 1-alpha (e.g., alpha=0.3 gives about 66-69% rather than 70%; alpha=0.1 gives 85-89% rather than 90%), and no standard deviations or confidence intervals are reported, so it is unclear whether deviations are statistically significant. Since the headline claim is a specified coverage probability, the authors should report the distribution of COV over runs and/or a coverage calibration plot, and explain the systematic under-coverage.
  4. [Eq. (6)] The threshold definition uses Ex in C I{s(pi(x))<tau} >= (|C|+1)/|C| (1-alpha), but no finite-sample quantile adjustment (e.g., ceil((|C|+1)(1-alpha))/|C|) is given, and the use of < versus <= is not specified. More importantly, the standard CP guarantee no longer applies once tau is adjusted by beta*rho in Eq. (8), so the paper should state clearly that Eq. (2) is an aspiration rather than a theorem, and provide empirical calibration evidence instead.
minor comments (5)
  1. [Eq. (10)] The cross-entropy loss is missing a summation over classes: it should be -sum_y gamma(x) * pi_tea(y|x) log pi_stu(y|x).
  2. [Fig. 2 and Fig. 3(b)] The legend in Fig. 2 lists 'TUI' instead of 'CUI'; Fig. 3(b) has an unlabeled x-axis, and the text does not state the range or units of the beta values shown.
  3. [Abstract and Sec. 5.2] The phrases 'plug-an-play' and 'play-and-plug' should be 'plug-and-play'.
  4. [Table 3] The table header contains the typo 'Soure Replay'; the storage amounts for the baseline are not clearly defined, making the comparison hard to interpret.
  5. [Algorithm 1] Step 3 says to compute nonconformity scores using Eq. (7), but the algorithm does not define the calibration labels or the score for a calibration pair (x_c, y_c); a concrete notation such as s(x,y) = 1 - pi_theta(y|x) should be used.

Circularity Check

1 steps flagged · score 6.0 of 10

Near-target coverage is controlled by an unreported scalar beta in Eq. (8), so the headline COV may be a tuned outcome rather than an independent prediction.

  1. fitted input called prediction [Sec. 4.1.3, Eq. (8); Sec. 5.3.4; Tables 1-2]
    "τˆ = τ ∗ − β · ρ, where β is a predefined factor. ... We also analysis the influence of different compensation factor β in Eq. (8), which represents the compensation level. The results are shown in Fig. 3(b), we find that small β decrease the compensation performance and large β may result in overcompensation."

    Eq. (8) makes the final threshold, and hence the reported coverage COV, a linear function of one scalar β: decreasing τˆ adds labels to every prediction set, so coverage is monotonically nondecreasing in β. The paper reports no fixed β, no selection rule, and no held-out tuning protocol; the only discussion treats β as a compensation knob with an optimum. On any given benchmark stream one can therefore choose β to push COV to, or near, 1 − α, so the 'near-expected coverage' in Tables 1-2 is a tuned consequence of Eq. (8), not an independent prediction of the CUI construction. This concern is compounded because ρ in Eq. (5) is an unnormalized sum over |C| × |B|, so even a fixed β cannot transfer across the different calibration and batch sizes used in the three benchmarks.

full rationale

The paper's conformal machinery is otherwise largely self-contained: τ ∗ comes from a standard quantile over a labeled calibration set, the prediction set is formed by thresholding nonconformity scores, and the use of JS divergence between joint source/current-model representations is a concrete, externally checkable domain-shift heuristic. No load-bearing self-citation chain or uniqueness theorem is invoked; the authors' prior CTTA papers appear only as related work and baselines. The central circularity risk is the coverage claim itself. Because Eq. (8) introduces an unreported free parameter β that directly and monotonically controls the prediction-set threshold, and because the paper's main CP validation metric is how close COV is to 1 − α, the reported near-target coverage cannot be distinguished from a per-benchmark calibration outcome. The sensitivity discussion in Sec. 5.3.4 explicitly frames β as a tuning knob, and no fixed value or reproducible selection rule is supplied. Under the paper's own equations, the coverage result is therefore partly constructed by the choice of β rather than being an independent first-principles prediction. The adaptation improvements are interesting, but they inherit this uncertainty because the CPAda weights in Eq. (11) are computed from the same β-dependent prediction sets. This warrants a partial-circularity score of 6.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on one under-specified hyperparameter (beta) and one unproven structural assumption (linear relation between rho and the coverage gap). The method does not invent new physical or mathematical entities, but it does rely on a static labeled calibration set and on the current model remaining useful for score computation. The calibration set size is a hand-chosen parameter that affects the stability of the quantile. No fully self-contained or externally benchmarked derivation is given.

free parameters (2)
  • beta = not reported
    Eq. (8) uses tau_hat = tau* - beta * rho; Sec. 5.3.4 analyzes beta but Tables 1-3 do not state the value used. The coverage outcome is directly controlled by this factor.
  • calibration set size = 50 (CIFAR10-C), 100 (CIFAR100-C), 500 (ImageNet-C)
    Chosen by hand in Sec. 5.1; affects quantile stability and CP performance but is not derived from any principle.
assumptions (4)
  • domain assumption The calibration set is exchangeable with the source distribution and remains a valid reference after model adaptation.
    Sec. 4.1.1 builds the CP quantile from a static calibration set; if the source calibration becomes unrepresentative under drift, the threshold is meaningless.
  • ad hoc to paper The JS divergence between joint source/current-model representations is monotonically and linearly related to the coverage gap.
    Eqs. (5) and (8) assume tau_hat = tau* - beta*rho; no theorem, bound, or monotonicity result is provided, and beta must be tuned.
  • ad hoc to paper Nonconformity scores for calibration can be computed with the current adapted model and remain comparable to source-model scores.
    Eq. (7) computes calibration scores with theta_crt; if the adapted model has drifted, the calibration quantile is computed on a different distribution than the original source model.
  • standard math Standard conformal prediction assumptions hold for the initial threshold before compensation.
    Sec. 4.1.3 inherits the split-conformal quantile construction from Vovk et al. and Sadinle et al.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Conformal Uncertainty Indicator for Continual Test-Time Adaptation." pith.science (2026). https://pith.science/paper/X7WC2CDD

@misc{pith2026250202998,
  author       = {Pith},
  title        = {Pith review of: Conformal Uncertainty Indicator for Continual Test-Time Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X7WC2CDD}},
  note         = {Machine review of arXiv:2502.02998}
}
read the original abstract

Continual Test-Time Adaptation (CTTA) aims to adapt models to sequentially changing domains during testing, relying on pseudo-labels for self-adaptation. However, incorrect pseudo-labels can accumulate, leading to performance degradation. To address this, we propose a Conformal Uncertainty Indicator (CUI) for CTTA, leveraging Conformal Prediction (CP) to generate prediction sets that include the true label with a specified coverage probability. Since domain shifts can lower the coverage than expected, making CP unreliable, we dynamically compensate for the coverage by measuring both domain and data differences. Reliable pseudo-labels from CP are then selectively utilized to enhance adaptation. Experiments confirm that CUI effectively estimates uncertainty and improves adaptation performance across various existing CTTA methods.

Figures

Figures reproduced from arXiv: 2502.02998 by the authors.

Figure 1
Figure 1. In the task of CTTA, a test sample x may be drawn from a different distribution in a long-term testing phase. Traditional methods rely on the self-adaptation based on the prediction and ignore the uncertainty may cause error accumulation. CUI provides a technique of uncertainty measurement based on CP. For the test sample, if CUI outputs a prediction set with small sizes (> 0), it is regarded as reliable and yields … view at source ↗
Figure 2
Figure 2. Visualization of coverage and inefficiency changes. [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Hyperparameter analysis on CIFAR100-to-CIFAR100C. [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Time and memory cost on CIFAR100-to-CIFAR100C. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 33 canonical work pages

  1. [1]

    Conformal prediction beyond exchangeability

    Rina Foygel Barber, Emmanuel J Candes, Aaditya Ramdas, and Ryan J Tibshirani. Conformal prediction beyond exchangeability. The Annals of Statistics, 51(2):816–845, 2023

    work page 2023

  2. [2]

    Improved online conformal prediction via strongly adaptive online learning

    Aadyot Bhatnagar, Huan Wang, Caiming Xiong, and Yu Bai. Improved online conformal prediction via strongly adaptive online learning. In International Conference on Machine Learning, pages 2337–2363, 2023

    work page 2023

  3. [3]

    A probabilistic framework for lifelong test-time adaptation

    Dhanajit Brahma and Piyush Rai. A probabilistic framework for lifelong test-time adaptation. In Proceedings of the Computer Vision and Pattern Recognition, 2023

    work page 2023

  4. [4]

    Intelligible models for healthcare: Predicting pneumonia risk and hospital 30-day readmission

    Rich Caruana, Yin Lou, Johannes Gehrke, Paul Koch, Marc Sturm, and Noemie Elhadad. Intelligible models for healthcare: Predicting pneumonia risk and hospital 30-day readmission. In Proceedings of the ACM SIGKDD international conference on knowledge discovery and data mining, pages 1721–1730, 2015

    work page 2015

  5. [5]

    SATA: Source Anchoring and Target Alignment Network for Continual Test Time Adaptation

    Goirik Chakrabarty, Manogna Sreenivas, and Soma Biswas. Sata: Source anchoring and target alignment network for continual test time adaptation. arXiv preprint arXiv:2304.10113, 2023

    work page Pith review arXiv 2023

  6. [6]

    A simple signal for domain shift

    Goirik Chakrabarty, Manogna Sreenivas, and Soma Biswas. A simple signal for domain shift. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3577–3584, 2023. 10 Conformal Uncertainty Indicator for Continual Test-Time Adaptation

    work page 2023

  7. [7]

    Each test image deserves a specific prompt: Continual test-time adaptation for 2d medical image segmentation

    Ziyang Chen, Yongsheng Pan, Yiwen Ye, Mengkang Lu, and Yong Xia. Each test image deserves a specific prompt: Continual test-time adaptation for 2d medical image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11184–11193, 2024

    work page 2024

  8. [8]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2009

    work page 2009

Show all 42 references
  1. [9]

    Robust mean teacher for continual and gradual test-time adaptation

    Mario Döbler, Robert A Marsden, and Bin Yang. Robust mean teacher for continual and gradual test-time adaptation. In Proceedings of the Computer Vision and Pattern Recognition, 2023

    2023

  2. [10]

    Non-exchangeable conformal risk control

    António Farinhas, Chrysoula Zerva, Dennis Ulmer, and André FT Martins. Non-exchangeable conformal risk control. arXiv preprint arXiv:2310.01262, 2023

    2023 arXiv

  3. [11]

    Plug-in martingales for testing exchangeability on-line

    Valentina Fedorova, Alex Gammerman, Ilia Nouretdinov, and Vladimir V ovk. Plug-in martingales for testing exchangeability on-line. In Proceedings of the International Conference on International Conference on Machine Learning, pages 923–930, 2012

    2012

  4. [12]

    Dropout as a bayesian approximation: Representing model uncertainty in deep learning

    Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In International conference on Machine Ltoearning, pages 1050–1059. PMLR, 2016

    2016

  5. [13]

    Conformal inference for online prediction with arbitrary distribution shifts

    Isaac Gibbs and Emmanuel Candès. Conformal inference for online prediction with arbitrary distribution shifts. arXiv preprint arXiv:2208.08401, 2022

    2022 arXiv

  6. [14]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the Computer Vision and Pattern Recognition, 2016

    2016

  7. [15]

    Online domain adaptation of a pre-trained cascade of classifiers

    Vidit Jain and Erik Learned-Miller. Online domain adaptation of a pre-trained cascade of classifiers. InProceedings of the Computer Vision and Pattern Recognition, 2011

    2011

  8. [16]

    Conformal decision theory: Safe autonomous decisions from imperfect predictions

    Jordan Lekeufack, Anastasios A Angelopoulos, Andrea Bajcsy, Michael I Jordan, and Jitendra Malik. Conformal decision theory: Safe autonomous decisions from imperfect predictions. arXiv preprint arXiv:2310.05921, 2023

    2023 arXiv

  9. [17]

    Variational continual test-time adaptation

    Fan Lyu, Kaile Du, Yuyang Li, Hanyu Zhao, Zhang Zhang, Guangcan Liu, and Liang Wang. Variational continual test-time adaptation. arXiv preprint arXiv:2402.08182, 2024

    2024

  10. [18]

    Multi-domain multi-task rehearsal for lifelong learning

    Fan Lyu, Shuai Wang, Wei Feng, Zihan Ye, Fuyuan Hu, and Song Wang. Multi-domain multi-task rehearsal for lifelong learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 8819–8827, 2021

    2021

  11. [19]

    A simple baseline for bayesian uncertainty in deep learning

    Wesley J Maddox, Pavel Izmailov, Timur Garipov, Dmitry P Vetrov, and Andrew Gordon Wilson. A simple baseline for bayesian uncertainty in deep learning. In Advances in neural information processing systems , volume 32, 2019

    2019

  12. [20]

    Rdumb: A simple approach that questions our progress in continual test-time adaptation

    Ori Press, Steffen Schneider, Matthias Kümmerer, and Matthias Bethge. Rdumb: A simple approach that questions our progress in continual test-time adaptation. Advances in Neural Information Processing Systems, 36, 2024

    2024

  13. [21]

    Post-hoc calibration of neural networks

    Amir Rahimi, Kartik Gupta, Thalaiyasingam Ajanthan, Thomas Mensink, Cristian Sminchisescu, and Richard Hartley. Post-hoc calibration of neural networks. arXiv preprint arXiv:2006.12807, 2, 2020

    2006 arXiv

  14. [22]

    Experience replay for continual learning

    David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory Wayne. Experience replay for continual learning. Advances in neural information processing systems, 32, 2019

    2019

  15. [23]

    Least ambiguous set-valued classifiers with bounded error levels

    Mauricio Sadinle, Jing Lei, and Larry Wasserman. Least ambiguous set-valued classifiers with bounded error levels. Journal of the American Statistical Association, 114(525):223–234, 2019

    2019

  16. [24]

    Controllable continual test-time adaptation

    Ziqi Shi, Fan Lyu, Ye Liu, Fanhua Shang, Fuyuan Hu, Wei Feng, Zhang Zhang, and Liang Wang. Controllable continual test-time adaptation. arXiv preprint arXiv:2405.14602, 2024

    2024 arXiv

  17. [25]

    Ar-tta: A simple method for real-world continual test-time adaptation

    Damian Sójka, Sebastian Cygert, Bartłomiej Twardowski, and Tomasz Trzci´nski. Ar-tta: A simple method for real-world continual test-time adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3491–3495, 2023

    2023

  18. [26]

    Exploring example influence in continual learning

    Qing Sun, Fan Lyu, Fanhua Shang, Wei Feng, and Liang Wan. Exploring example influence in continual learning. Advances in Neural Information Processing Systems, 35:27075–27086, 2022

    2022

  19. [27]

    Test-time training with self-supervision for generalization under distribution shifts

    Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International Conference on Machine Learning, 2020

    2020

  20. [28]

    Less is more: Pseudo-label filtering for continual test-time adaptation

    Jiayao Tan, Fan Lyu, Chenggong Ni, Tingliang Feng, Fuyuan Hu, Zhang Zhang, Shaochuang Zhao, and Liang Wang. Less is more: Pseudo-label filtering for continual test-time adaptation. arXiv preprint arXiv:2406.02609, 2024. 11 Conformal Uncertainty Indicator for Continual Test-Tim...

    2024 arXiv

  21. [29]

    Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results

    Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In Proceedings of the Advances in Neural Information Processing Systems, 2017

    2017

  22. [30]

    Post-hoc uncertainty calibration for domain drift scenarios

    Christian Tomani, Sebastian Gruber, Muhammed Ebrar Erdem, Daniel Cremers, and Florian Buettner. Post-hoc uncertainty calibration for domain drift scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10124–10132, 2021

    2021

  23. [31]

    Brain-inspired replay for continual learning with artificial neural networks

    Gido M Van de Ven, Hava T Siegelmann, and Andreas S Tolias. Brain-inspired replay for continual learning with artificial neural networks. Nature communications, 11(1):4069, 2020

    2020

  24. [32]

    Inductive conformal martingales for change-point detection

    Denis V olkhonskiy, Evgeny Burnaev, Ilia Nouretdinov, Alexander Gammerman, and Vladimir V ovk. Inductive conformal martingales for change-point detection. In Conformal and Probabilistic Prediction and Applications, pages 132–153, 2017

    2017

  25. [33]

    Springer, 2005

    Vladimir V ovk, Alexander Gammerman, and Glenn Shafer.Algorithmic learning in a random world, volume 29. Springer, 2005

    2005

  26. [34]

    Computationally efficient versions of conformal predictive distributions

    Vladimir V ovk, Ivan Petej, Ilia Nouretdinov, Valery Manokhin, and Alexander Gammerman. Computationally efficient versions of conformal predictive distributions. Neurocomputing, 397:292–308, 2020

    2020

  27. [35]

    Tent: Fully test-time adaptation by entropy minimization

    Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In Proceedings of the International Conference on Learning Representations, 2020

    2020

  28. [36]

    Continual test-time domain adaptation

    Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of the Computer Vision and Pattern Recognition, 2022

    2022

  29. [37]

    Aggregated residual transformations for deep neural networks

    Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the Computer Vision and Pattern Recognition, 2017

    2017

  30. [38]

    Exploring safety supervision for continual test-time domain adaptation

    Xu Yang, Yanan Gu, Kun Wei, and Cheng Deng. Exploring safety supervision for continual test-time domain adaptation. In Proceedings of the International Joint Conference on Artificial Intelligence, 2023

    2023

  31. [39]

    Test-time recalibration of conformal predictors under distribution shift based on unlabeled examples

    Fatih Furkan Yilmaz and Reinhard Heckel. Test-time recalibration of conformal predictors under distribution shift based on unlabeled examples. arXiv preprint arXiv:2210.04166, 2022

    2022 arXiv

  32. [40]

    Adaptive conformal predictions for time series

    Margaux Zaffran, Olivier Féron, Yannig Goude, Julie Josse, and Aymeric Dieuleveut. Adaptive conformal predictions for time series. In International Conference on Machine Learning, pages 25834–25866, 2022

    2022

  33. [41]

    Wide residual networks

    Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In Procedings of the British Machine Vision Conference, 2016

    2016

  34. [42]

    Coverage-guaranteed prediction sets for out-of-distribution data

    Xin Zou and Weiwei Liu. Coverage-guaranteed prediction sets for out-of-distribution data. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 17263–17270, 2024. 12

    2024

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.