Pith. sign in

REVIEW 5 major objections 6 minor 61 references

TTVD: Towards a Geometric Framework for Test-Time Adaptation Based on Voronoi Diagram

T0 review · 5 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

Pith's one-line read The paper claims that neighbor-based test-time adaptation is secretly a Voronoi diagram, and that upgrading it with cluster-induced influence and weighted cell boundaries yields the lowest errors and best calibration on four corruption…

desk verdict A fresh geometric take on TTA, but the algorithm as written has a sign inversion and an ill-defined exponent that make the headline results unreproducible. read the letter →

arxiv 2412.07980 v1 pith:UD3GI2X3 submitted 2024-12-10 cs.CV cs.AI

classification cs.CVcs.AI
keywords test-timeadaptationVoronoidiagramcluster-inducedPowerdistributionshiftentropyminimizationmodelcalibrationnoisysamplefiltering
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

Test-time adaptation fixes a model's drift when the test distribution shifts, but existing methods often rely on either self-supervision or entropy minimization, which can overfit or degrade over time. This paper argues that the neighbor-based family of methods is secretly a Voronoi diagram: the feature space is divided into cells around class prototypes, and adaptation pushes test features into the right cells. Building on that geometric reading, the paper constructs TTVD, which combines a cluster-induced Voronoi diagram (sites are clusters, so multiple rotations of a sample jointly vote) with a power diagram (cells carry weights, so noisy samples near boundaries can be filtered). Under the standardized TTAB evaluation, TTVD reports the lowest classification error and expected calibration error on CIFAR-10-C, CIFAR-100-C, ImageNet-C, and ImageNet-R, and it keeps its edge at small batch sizes and under label shift. If the geometric reading is right, it gives a single lens for understanding and improving a broad class of adaptation algorithms.

What carries the argument

The load-bearing object is the Voronoi diagram used as an adaptation loss, not merely as a classifier. A feature $z$ receives soft label $\beta(-d(z, \mu_k)/\tau)$ over the class sites $\mu_k$, and the training signal is the entropy of that soft-label vector, so adaptation is equivalent to pulling $z$ into the cell of the correct prototype. The two refinements are defined by changing the distance: CIVD aggregates influences from a cluster of sites via $F(z, C_k) = -\mathrm{sign}(\gamma) \sum_\alpha d(\mu_k^{(\alpha)}, z)^\gamma$, and the power diagram replaces $d^2$ with $d^2 - v_k^2$, adding a per-cell weight. The weighted boundaries of the power diagram are used for noisy-sample filtering by subtracting the PD partition from the VD partition, and because CIVD and PD are parallel structures, they compose into the cluster-induced power diagram (CIPD) that TTVD actually runs at test time.

What would settle it

Run TTVD on CIFAR-10-C with $\gamma$ swept over a grid (e.g., $-0.2$, $-0.5$, $-0.8$, $-1.5$) and with the CIVD influence function replaced by an ordinary distance average ($\gamma = 1$), keeping all other settings fixed: if the error advantage over Tent and SAR concentrates at the single published value $\gamma = -0.8$ and collapses at neighbouring values, or if the plain-distance version keeps the same gain, then the specific geometric influence term is not the cause of the reported improvement.

Watch

Extended reading notes

Core claim

The paper's central claim is that any neighbor-based TTA method—one that compares test features to stored prototypes—is geometrically a Voronoi diagram, and that the diagram's two generalizations solve the two known failure modes of those methods. The first generalization, the cluster-induced Voronoi diagram (CIVD), replaces each single prototype by a cluster of prototypes obtained from self-supervised rotations, and defines the cell assignment by a joint influence function $F(z, C_k) = -\mathrm{sign}(\gamma) \sum_\alpha d(\mu_k^{(\alpha)}, z)^\gamma$ with $\gamma = -0.8$. This makes the soft label a pooled vote over all rotated views, which the paper shows corrects misclassifications that any single view would make. The second generalization, the power diagram (PD), gives every cell a weight so the boundary can shift; subtracting the PD from the plain Voronoi diagram isolates a region of high-gradient, unreliable samples, and those samples are excluded from adaptation. The paper reports that the combined method, TTVD, achieves the best error and expected calibration error in the TTAB benchmark on four datasets, and that the CIVD step contributes the largest single improvement (5.7% average error reduction on CIFAR-10-C) before the PD step adds another 2.2%.

Load-bearing premise

The method's gains rest on the hand-chosen influence function $F(z, C_k) = -\mathrm{sign}(\gamma) \sum_\alpha d(\mu_k^{(\alpha)}, z)^\gamma$ with $\gamma = -0.8$ and temperature $\tau = 1$; the paper gives no derivation of this choice, and if that soft-label signal is not a reliable training target across the full range of corruptions, the reported improvements would shrink or disappear.

Editorial extensions

If this is right

  • A single Voronoi-style loss can carry both self-supervision and entropy minimization without tuning a trade-off weight, because the cluster-induced influence function combines the two objectives in one mechanism.
  • The power diagram's boundary subtraction filters unreliable samples without a separate entropy threshold, which the paper shows is error-prone on the entropy landscape.
  • TTVD keeps most of its advantage at batch sizes 32, 16, and 8 and under Dirichlet label shift with $\alpha = 0.01$, meaning the method is usable in the non-i.i.d. online streams where TTA is normally deployed.
  • Adaptation curves on ImageNet-C noise show TTVD still descending at the last batch, while Tent and SAR plateau, which the paper interprets as a reduced tendency to overfit to the evolving test distribution.
  • The method is insensitive to the precision of the precomputed class means: using 10%, 5%, or 1% of ImageNet gives essentially identical error, so the offline site computation can be made cheap at scale.

Reading between the lines

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

  • The Voronoi reading suggests that any prototype-based TTA loss is a choice of influence kernel; replacing the power-law distance with a Gaussian kernel would produce a family of methods and could reveal whether the reported gains are specific to $\gamma = -0.8$ or general to kernel-based cell assignment.
  • Since the power diagram is used only to filter samples, a natural variant would use the same weighted boundaries to down-weight rather than discard boundary samples, preserving gradient information from shifted but informative features.
  • The paper's geometric framing implies that adaptation methods can be compared by their cell geometry rather than only by accuracy; a testable extension is to predict which corruption types a method will fail on from the shape of its Voronoi cells in a 2D embedding.
  • The paper's own footnote warns that oracle-model-selection numbers overstate real-world performance; a fair replication should use the non-subscripted, label-free numbers when comparing to deployed settings.
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

5 major / 6 minor

Summary. The paper proposes TTVD, a test-time adaptation method framed through computational geometry. It first observes that neighbor-based TTA methods correspond to a Voronoi Diagram, then generalizes to a Cluster-induced Voronoi Diagram (CIVD) whose influence function aggregates multiple sites per class, and further to a Cluster-induced Power Diagram (CIPD) with per-site weights for boundary flexibility and noisy-sample filtering. The method is evaluated with the TTAB toolkit on CIFAR-10-C, CIFAR-100-C, ImageNet-C, and ImageNet-R, reporting the lowest classification error and expected calibration error among the compared methods, plus additional experiments on batch size, label shift, and robustness of the class-mean estimates. The central claim is that the geometric formulation unifies self-supervision and entropy minimization while avoiding negative transfer, and that CIPD's flexible boundaries improve subsample selection.

Significance. If the claims held, TTVD would be a meaningful contribution to TTA: it is evaluated on an externally maintained benchmark (TTAB), it compares against a broad set of baselines, and it reports consistent improvements in both error and calibration across four benchmarks. The paper also shows robustness to reduced precision in the class means, which is practically useful. However, the significance is currently undermined by specifying inconsistencies that make the published algorithm not reproducible from the text: the sign flip in the soft-label computation, the undefined fractional power in Eq. (6), and the absence of the filtering mechanism that is central to the CIPD variant. These do not refute the geometric idea, but they prevent the reader from verifying the reported results and need to be corrected and supplemented with code or precise algorithmic details.

major comments (5)
  1. [§3.2, Eq. (4), Eq. (6), Algorithms 2–3 (Appendix H)] The soft-label computation in Algorithms 2 and 3 inverts the influence function defined in Definitions 3.2 and 3.4. With γ = −0.8, F(z,C_k) = Σ_α d(μ_k^(α),z)^{−0.8} because −sign(γ)=+1, so the cell assignment r(z)=arg max_k F selects the cluster with the smallest aggregate distance, as intended. However, the algorithms compute y~_k = β(−F(z,C_k)+ε; τ), so the logit for the assigned (nearest) cluster is the most negative and its soft-label probability tends to 0 as d→0. Consequently, the entropy loss L_VD pushes features away from the very site that the partition rule assigns them to, contradicting the mechanism described in §3.1–§3.2 and the sample analysis in Appendix A.1, which relies on F without the minus sign. For γ>0 the same inversion occurs because −sign(γ)Σd^γ is minimized at the nearest site. The equations and the algorithm blocks cannot both be correct; either the minus sign in Algorithms 2/3 is a typo (it should be +F) or the reported results do not come from the specified algorithm. This must be corrected and verified by released code.
  2. [Definition 3.4, Eq. (6)] The Cluster-induced Power Diagram influence F(z,C_k) = −sign(γ) Σ_α {d(μ_k^(α),z)^2 − v_k^2}^γ is not well-defined for the announced γ = −0.8. The quantity d^2 − v^2 is negative for points inside the power circle of site k, and the fractional power of a negative number is complex; it is also singular at d^2 = v^2. The paper gives no clipping, absolute-value, sign-preserving power, or restriction on the domain, so the CIPD loss and its gradients are undefined over a substantial part of feature space. Because Table 2 credits CIPD with a further 2.2% improvement over CIVD, the evaluation of this core variant is not reproducible from the text. Please specify the exact definition used in practice (e.g., replacing the base by |d^2 − v^2| with sign, or clipping at a small positive value) and justify it geometrically.
  3. [§3.3, Algorithm 3, Table 2] The sample-filtering mechanism attributed to the Power Diagram is not implemented in the algorithm that is claimed to define TTVD. §3.3 describes subtracting the VD from the PD to detect noisy samples near boundaries, but Algorithm 3 only computes soft labels via Eq. (6) and applies the entropy loss; there is no step that excludes samples, no threshold, and no description of how the diagram subtraction is performed in feature space. Moreover, the PD weights v_k are never specified: Lemma 3.1 connects logistic-regression parameters to a power diagram, but it is unclear whether v_k are taken from the classifier biases, are learned, or are set to constants. Without these details, the 2.2% CIPD improvement over CIVD in Table 2 cannot be attributed to the proposed filtering, and the method cannot be re-implemented.
  4. [§4.2, Appendix D] The influence exponent γ = −0.8 and temperature τ = 1 are fixed hyperparameters, but no sensitivity analysis, ablation, or derivation is provided. The behavior of the influence function depends critically on the sign and magnitude of γ; for instance, sign(γ) determines whether the influence is attractive or repulsive, and the power-law exponent determines how quickly distant sites are discounted. Since TTVD's claims rest on this choice and the value is not selected by a principled criterion, the paper should report performance over a range of γ (and τ) and justify the chosen value. This is particularly important because the reported gains over strong baselines are modest (0.7–1.6% error), and the optimal γ may vary across datasets.
  5. [§4.2, Tables 1, 3, 5–12] No error bars or repeated-run statistics are reported for any of the results. The claimed improvements are small on several benchmarks (e.g., 0.7% on CIFAR-100-C and ImageNet-R), and TTA results are sensitive to the order of the online stream, batch composition, and random seeds. Without confidence intervals or at least multiple seeds, the statistical significance of the main empirical claim cannot be assessed. Please report mean and standard deviation over at least three runs, or otherwise justify that the reported differences are reproducible.
minor comments (6)
  1. [Throughout] There are numerous typographical and formatting issues: 'V oronoi' with a stray space appears repeatedly, 'inherit limitation' should be 'inherent limitation', 'iterpretability' should be 'interpretability', 'Internetional' should be 'International' (Appendix F), and 'Congugate' should be 'Conjugate' (Appendix F header). A copy-edit pass is needed.
  2. [Lemma 3.1] Lemma 3.1 is attributed to references [36,37] (DeepVoro and iVoro), but the logistic-regression power-diagram identity originates in the analysis of deep networks by Balestriero et al. (reference [5] in the bibliography). Please correct the citation.
  3. [References] References [21] and [22] are the same publication (Influence-based Voronoi diagrams of clusters, Computational Geometry 96:101746, 2021) and are listed twice. In addition, two non-identical entries for the pyvoro software [38,45] appear. Please deduplicate.
  4. [§3.2] The claim that CIVD 'avoids the negative transfer' is not directly supported by the experiments; no measurement or comparison of negative transfer is reported. Please either soften the claim or provide evidence.
  5. [Figure 1, Appendix C] The caption of Figure 1 refers to an 'Augmented Voronoi Diagram (by self-supervision)' while the text calls this CIVD. Align the terminology between the figure and the main text.
  6. [Appendix A.1] The sample-level analysis in Appendix A.1 shows only three hand-picked examples. It would be strengthened by reporting aggregate statistics over the misclassified set (e.g., what fraction of CIVD corrections are explained by the rotated views agreeing with the ground-truth class).

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: TTVD is validated on external TTAB benchmarks; self-cited CIVD definitions and power-diagram lemma are not load-bearing.

full rationale

The paper's central claim is an accuracy/ECE comparison on external test suites (CIFAR-10-C, CIFAR-100-C, ImageNet-C, ImageNet-R) run under the TTAB harness. No parameter is fitted to test labels and then renamed as a prediction; the final losses are fixed functions of distances to precomputed class-mean sites. The CIVD definition (Definition 3.2) and the power-diagram lemma (Lemma 3.1) are imported from the authors' prior computational-geometry work, but they are mathematical definitions and known facts used as building blocks, not results whose truth depends on accepting this paper's conclusions. The choice gamma=-0.8 and tau=1 (Appendix D) is a hand-set hyperparameter; it is under-justified but does not create a fitted-input-called-prediction loop because the same fixed setting is then evaluated on unseen test streams. The skeptic's observed sign inconsistency (Algorithms 2 and 3 use beta(-F) while Definitions 3.2 and 3.4 maximize F) and the undefined {d^2-v^2}^{-0.8} for d^2<v^2 are correctness and reproducibility defects, not circular reductions: they do not make Equations 3-6 equal to their own inputs by construction. Therefore no step satisfies the circularity bar; the self-citations are present but non-load-bearing, so the score is 2.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim rests on a geometric reformulation of TTA as distance-based soft-label entropy minimization. The only method-specific hand-tuned number is gamma (and a fixed tau), while the core assumptions are the validity of the influence function and the entropy objective. No new physical or mathematical entities are introduced.

free parameters (2)
  • gamma (influence exponent) = -0.8
    Controls the influence of distant sites in the CIVD/CIPD soft-label computation (Eq. 4 and Eq. 6). Chosen by hand (Appendix D) with no sensitivity analysis, likely tuned for benchmark performance.
  • tau (softmax temperature) = 1
    Temperature in softmax for soft labels in Eq. 3. Set to the standard default, not tuned.
assumptions (3)
  • standard math Logistic regression partitions feature space into a power diagram with v_k^2 = b_k + 1/4 ||W_k||^2 (Lemma 3.1).
    Adopted from Balestriero et al. and used to justify the power-diagram weights in Eq. 6.
  • domain assumption The cluster-induced Voronoi influence function F(z, C_k) = -sign(gamma) * sum_alpha d(mu_k^alpha, z)^gamma is a meaningful distance-to-cluster measure for soft labels.
    Imported from the authors' prior CIVD work and assumed to be beneficial for TTA without a dedicated derivation.
  • domain assumption Entropy of the distance-based soft-label distribution is an appropriate training signal for online test-time adaptation.
    The paper builds the adaptation loss as entropy of the soft labels but does not prove why this entropy is preferable to other objectives; this is the core modeling premise.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TTVD: Towards a Geometric Framework for Test-Time Adaptation Based on Voronoi Diagram." pith.science (2026). https://pith.science/paper/UD3GI2X3

@misc{pith2026241207980,
  author       = {Pith},
  title        = {Pith review of: TTVD: Towards a Geometric Framework for Test-Time Adaptation Based on Voronoi Diagram},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UD3GI2X3}},
  note         = {Machine review of arXiv:2412.07980}
}
read the original abstract

Deep learning models often struggle with generalization when deploying on real-world data, due to the common distributional shift to the training data. Test-time adaptation (TTA) is an emerging scheme used at inference time to address this issue. In TTA, models are adapted online at the same time when making predictions to test data. Neighbor-based approaches have gained attention recently, where prototype embeddings provide location information to alleviate the feature shift between training and testing data. However, due to their inherit limitation of simplicity, they often struggle to learn useful patterns and encounter performance degradation. To confront this challenge, we study the TTA problem from a geometric point of view. We first reveal that the underlying structure of neighbor-based methods aligns with the Voronoi Diagram, a classical computational geometry model for space partitioning. Building on this observation, we propose the Test-Time adjustment by Voronoi Diagram guidance (TTVD), a novel framework that leverages the benefits of this geometric property. Specifically, we explore two key structures: 1) Cluster-induced Voronoi Diagram (CIVD): This integrates the joint contribution of self-supervision and entropy-based methods to provide richer information. 2) Power Diagram (PD): A generalized version of the Voronoi Diagram that refines partitions by assigning weights to each Voronoi cell. Our experiments under rigid, peer-reviewed settings on CIFAR-10-C, CIFAR-100-C, ImageNet-C, and ImageNet-R shows that TTVD achieves remarkable improvements compared to state-of-the-art methods. Moreover, extensive experimental results also explore the effects of batch size and class imbalance, which are two scenarios commonly encountered in real-world applications. These analyses further validate the robustness and adaptability of our proposed framework.

Figures

Figures reproduced from arXiv: 2412.07980 by the authors.

Figure 1
Figure 1. (A) Visualization of space partitions induced by Voronoi Diagram, Power [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Noisy sample filtering by diagram subtraction. (a) Entropy landscape of [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. (Left) Illustrations on differences between VD, CIVD and CIPD. (Right) [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Comparison on the Adaptation Curves on different noise perturbations in [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: A misclassified “bike” sample corrected by CIVD. The x-axis denotes the [PITH_FULL_IMAGE:figures/full_fig_p019_5.png]
Figure 6
Figure 6. Figure 6: A misclassified “bus” sample corrected by CIVD. The x-axis denotes the [PITH_FULL_IMAGE:figures/full_fig_p020_6.png]
Figure 7
Figure 7. Figure 7: A misclassified “clock” sample corrected by CIVD. The x-axis denotes the [PITH_FULL_IMAGE:figures/full_fig_p021_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

61 extracted references · 48 canonical work pages

  1. [1]

    Self-supervised learning for domain adaptation on point clouds

    Idan Achituve, Haggai Maron, and Gal Chechik. Self-supervised learning for domain adaptation on point clouds. In Proceedings of the IEEE/CVF winter con- ference on applications of computer vision, pages 123–133, 2021

  2. [2]

    Aurenhammer

    F. Aurenhammer. Power diagrams: Properties, algorithms and applications.SIAM Journal on Computing, 16(1):78–96, 1987

  3. [3]

    V oronoi diagrams—a survey of a fundamental geometric data structure

    Franz Aurenhammer. V oronoi diagrams—a survey of a fundamental geometric data structure. ACM Comput. Surv., 23(3):345–405, sep 1991

  4. [4]

    Baraniuk

    Randall Balestriero and Richard G. Baraniuk. Mad max: Affine spline insights into deep learning. Proceedings of the IEEE, 109(5):704–727, 2021

  5. [5]

    The geometry of deep networks: Power diagram subdivision

    Randall Balestriero, Romain Cosentino, Behnaam Aazhang, and Richard Bara- niuk. The geometry of deep networks: Power diagram subdivision. Advances in Neural Information Processing Systems, 32, 2019

  6. [6]

    Characterizing large language model geometry solves toxicity detection and generation

    Randall Balestriero, Romain Cosentino, and Sarath Shekkizhar. Characterizing large language model geometry solves toxicity detection and generation. arXiv preprint arXiv:2312.01648, 2023

  7. [7]

    Chen, Ziyun Huang, Yangwei Liu, and Jinhui Xu

    Danny Z. Chen, Ziyun Huang, Yangwei Liu, and Jinhui Xu. On clustering in- duced voronoi diagrams. In 2013 IEEE 54th Annual Symposium on Foundations of Computer Science, pages 390–399, 2013

  8. [8]

    Chen, Ziyun Huang, Yangwei Liu, and Jinhui Xu

    Danny Z. Chen, Ziyun Huang, Yangwei Liu, and Jinhui Xu. On clustering in- duced voronoi diagrams. SIAM Journal on Computing, 46(6):1679–1711, 2017. 13

Show all 61 references
  1. [9]

    Ac- tion segmentation with joint self-supervised temporal domain adaptation

    Min-Hung Chen, Baopu Li, Yingze Bao, Ghassan AlRegib, and Zsolt Kira. Ac- tion segmentation with joint self-supervised temporal domain adaptation. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR), June 2020

  2. [10]

    Self-ensembling for vi- sual domain adaptation

    Geoffrey French, Michal Mackiewicz, and Mark Fisher. Self-ensembling for vi- sual domain adaptation. arXiv preprint arXiv:1706.05208, 2017

  3. [11]

    Test-time training with masked autoencoders

    Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei Efros. Test-time training with masked autoencoders. Advances in Neural Information Processing Systems, 35:29374–29385, 2022

  4. [12]

    Domain- adversarial training of neural networks

    Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, Franc ¸ois Laviolette, Mario March, and Victor Lempitsky. Domain- adversarial training of neural networks. Journal of machine learning research , 17(59):1–35, 2016

  5. [13]

    NOTE: Robust continual test-time adaptation against temporal cor- relation

    Taesik Gong, Jongheon Jeong, Taewon Kim, Yewon Kim, Jinwoo Shin, and Sung-Ju Lee. NOTE: Robust continual test-time adaptation against temporal cor- relation. In Advances in Neural Information Processing Systems (NeurIPS), 2022

  6. [14]

    Generative adver- sarial nets

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde- Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adver- sarial nets. Advances in neural information processing systems, 27, 2014

  7. [15]

    Test-time adaptation via conjugate pseudo-labels

    Sachin Goyal, Mingjie Sun, Aditi Raghunanthan, and Zico Kolter. Test-time adaptation via conjugate pseudo-labels. Advances in Neural Information Pro- cessing Systems, 2022

  8. [16]

    Test-time training on nearest neighbors for large lan- guage models

    Moritz Hardt and Yu Sun. Test-time training on nearest neighbors for large lan- guage models. In The Twelfth International Conference on Learning Representa- tions, 2024

  9. [17]

    Deep residual learn- ing for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learn- ing for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

  10. [18]

    The many faces of robustness: A critical analysis of out-of-distribution generalization

    Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF international c...

  11. [19]

    Benchmarking neural network robust- ness to common corruptions and perturbations.arXiv preprint arXiv:1903.12261, 2019

    Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robust- ness to common corruptions and perturbations.arXiv preprint arXiv:1903.12261, 2019

  12. [20]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems , 33:6840–6851, 2020. 14

  13. [21]

    Chen, and Jinhui Xu

    Ziyun Huang, Danny Z. Chen, and Jinhui Xu. Influence-based voronoi diagrams of clusters. Computational Geometry, 96:101746, 2021

  14. [22]

    Influence-based voronoi diagrams of clusters

    Ziyun Huang, Danny Z Chen, and Jinhui Xu. Influence-based voronoi diagrams of clusters. Computational Geometry, 96:101746, 2021

  15. [23]

    Baraniuk

    Ahmed Imtiaz Humayun, Randall Balestriero, Guha Balakrishnan, and Richard G. Baraniuk. Splinecam: Exact visualization and characterization of deep network geometry and decision boundaries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CV...

  16. [24]

    Test-time classifier adjustment module for model-agnostic domain generalization

    Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. In M. Ranzato, A. Beygelzimer, Y . Dauphin, P.S. Liang, and J. Wortman Vaughan, editors,Advances in Neural In- formation Processing Systems, volume 34, pages 2427...

  17. [25]

    Test-time adap- tation via self-training with nearest neighbor information

    Minguk Jang, Sae-Young Chung, and Hye Won Chung. Test-time adap- tation via self-training with nearest neighbor information. arXiv preprint arXiv:2207.10792, 2022

  18. [26]

    Rotograd: Gradient homogenization in multi- task learning

    Adri ´an Javaloy and Isabel Valera. Rotograd: Gradient homogenization in multi- task learning. In International Conference on Learning Representations, 2022

  19. [27]

    Forkmerge: Mitigating negative transfer in auxiliary-task learning

    Junguang Jiang, Baixu Chen, Junwei Pan, Ximei Wang, Dapeng Liu, jie jiang, and Mingsheng Long. Forkmerge: Mitigating negative transfer in auxiliary-task learning. In Thirty-seventh Conference on Neural Information Processing Sys- tems, 2023

  20. [28]

    Deep adversarial attention alignment for unsupervised domain adaptation: the benefit of target expectation maximization

    Guoliang Kang, Liang Zheng, Yan Yan, and Yi Yang. Deep adversarial attention alignment for unsupervised domain adaptation: the benefit of target expectation maximization. In Proceedings of the European conference on computer vision (ECCV), pages 401–416, 2018

  21. [29]

    Venkatesh Babu

    Jogendra Nath Kundu, Naveen Venkat, Rahul M V , and R. Venkatesh Babu. Uni- versal source-free domain adaptation. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), June 2020

  22. [30]

    Self-supervised label augmenta- tion via input transformations

    Hankook Lee, Sung Ju Hwang, and Jinwoo Shin. Self-supervised label augmenta- tion via input transformations. InInternational Conference on Machine Learning, pages 5714–5724. PMLR, 2020

  23. [31]

    Domain generaliza- tion with adversarial feature learning

    Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex C Kot. Domain generaliza- tion with adversarial feature learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5400–5409, 2018

  24. [32]

    Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation

    Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In Interna- tional Conference on Machine Learning (ICML), pages 6028–6039, 2020. 15

  25. [33]

    Source data- absent unsupervised domain adaptation through hypothesis transfer and label- ing transfer

    Jian Liang, Dapeng Hu, Yunbo Wang, Ran He, and Jiashi Feng. Source data- absent unsupervised domain adaptation through hypothesis transfer and label- ing transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2021. In Press

  26. [34]

    Source-free domain adaptation for se- mantic segmentation

    Yuang Liu, Wei Zhang, and Jun Wang. Source-free domain adaptation for se- mantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1215–1224, June 2021

  27. [35]

    TTT++: When does self-supervised test- time training fail or thrive? In A

    Yuejiang Liu, Parth Kothari, Bastien Germain van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. TTT++: When does self-supervised test- time training fail or thrive? In A. Beygelzimer, Y . Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neura...

  28. [36]

    Few-shot learning as cluster-induced voronoi diagrams: A geometric approach, 2022

    Chunwei Ma, Ziyun Huang, Mingchen Gao, and Jinhui Xu. Few-shot learning as cluster-induced voronoi diagrams: A geometric approach, 2022

  29. [37]

    Progressive voronoi diagram subdivision enables accurate data-free class-incremental learning

    Chunwei Ma, Zhanghexuan Ji, Ziyun Huang, Yan Shen, Mingchen Gao, and Jinhui Xu. Progressive voronoi diagram subdivision enables accurate data-free class-incremental learning. In The Eleventh International Conference on Learn- ing Representations, 2023

  30. [38]

    Martin McBride. pyvoro. https://github.com/martinmcbride/ generativepy, 2014

  31. [39]

    Mnist-c: A robustness benchmark for computer vision

    Norman Mu and Justin Gilmer. Mnist-c: A robustness benchmark for computer vision. arXiv preprint arXiv:1906.02337, 2019

  32. [40]

    Efficient test-time model adaptation without forgetting

    Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In The Internetional Conference on Machine Learning, 2022

  33. [41]

    Towards stable test-time adaptation in dynamic wild world

    Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. In Internetional Conference on Learning Representations, 2023

  34. [42]

    Un- supervised intra-domain adaptation for semantic segmentation through self- supervision

    Fei Pan, Inkyu Shin, Francois Rameau, Seokju Lee, and In So Kweon. Un- supervised intra-domain adaptation for semantic segmentation through self- supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020

  35. [43]

    The en- tropy enigma: Success and failure of entropy minimization

    Ori Press, Ravid Shwartz-Ziv, Yann LeCun, and Matthias Bethge. The en- tropy enigma: Success and failure of entropy minimization. arXiv preprint arXiv:2405.05012, 2024

  36. [44]

    Improving robustness against common corruptions by covariate shift adaptation

    Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Bren- del, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Advances in neural information processing systems , 33:11539–11551, 2020. 16

  37. [45]

    Andrey Sobolev. pyvoro. https://github.com/joe-jordan/pyvoro, 2014

  38. [46]

    Yu Sun, Xiaolong Wang, Liu Zhuang, John Miller, Moritz Hardt, and Alexei A. Efros. Test-time training with self-supervision for generalization under distribu- tion shifts. In ICML, 2020

  39. [47]

    Sequence to sequence learning with neural networks

    Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks. Advances in neural information processing systems , 27, 2014

  40. [48]

    Discriminative adversarial domain adaptation

    Hui Tang and Kui Jia. Discriminative adversarial domain adaptation. In Proceed- ings of the AAAI conference on artificial intelligence , volume 34, pages 5940– 5947, 2020

  41. [49]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017

  42. [50]

    Nouvelles applications des param `etres continus `a la th ´eorie des formes quadratiques

    Georges V oronoi. Nouvelles applications des param `etres continus `a la th ´eorie des formes quadratiques. deuxi `eme m ´emoire. recherches sur les parall ´ello`edres primitifs. Journal f ¨ur die reine und angewandte Mathematik (Crelles Journal) , 1908(134):198–287, 1908

  43. [51]

    Nouvelles applications des param `etres continus `a la th ´eorie des formes quadratiques

    Georges V oronoi. Nouvelles applications des param `etres continus `a la th ´eorie des formes quadratiques. premier m ´emoire. sur quelques propri ´et´es des formes quadratiques positives parfaites. Journal f ¨ur die reine und angewandte Mathe- matik (Crelles Journal), 1908(13...

  44. [52]

    Tent: Fully test-time adaptation by entropy minimization

    Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Dar- rell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021

  45. [53]

    Continual test-time do- main adaptation

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

  46. [54]

    A MAX-AFFINE SPLINE PERSPECTIVE OF RECURRENT NEURAL NETWORKS

    Zichao Wang, Randall Balestriero, and Richard Baraniuk. A MAX-AFFINE SPLINE PERSPECTIVE OF RECURRENT NEURAL NETWORKS. In Inter- national Conference on Learning Representations, 2019

  47. [55]

    Exploiting the intrinsic neighborhood structure for source-free domain adaptation.Advances in neural information processing systems, 34:29393–29405, 2021

    Shiqi Yang, Joost van de Weijer, Luis Herranz, Shangling Jui, et al. Exploiting the intrinsic neighborhood structure for source-free domain adaptation.Advances in neural information processing systems, 34:29393–29405, 2021

  48. [56]

    Trust your good friends: Source-free domain adaptation by reciprocal neighborhood clustering

    Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, Shangling Jui, and Jian Yang. Trust your good friends: Source-free domain adaptation by reciprocal neighborhood clustering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(12):15883–15895, 2023. 17

  49. [57]

    Max-affine spline insights into deep network pruning.Transactions on Machine Learning Research, 2022

    Haoran You, Randall Balestriero, Zhihan Lu, Yutong Kou, Huihong Shi, Shun- yao Zhang, Shang Wu, Yingyan Lin, and Richard Baraniuk. Max-affine spline insights into deep network pruning.Transactions on Machine Learning Research, 2022

  50. [58]

    Wide residual networks.arXiv preprint arXiv:1605.07146, 2016

    Sergey Zagoruyko and Nikos Komodakis. Wide residual networks.arXiv preprint arXiv:1605.07146, 2016

  51. [59]

    Zhang, S

    M. Zhang, S. Levine, and C. Finn. MEMO: Test time robustness via adaptation and augmentation. 2021

  52. [60]

    AdaNPC: Exploring non-parametric classifier for test-time adaptation

    Yifan Zhang, Xue Wang, Kexin Jin, Kun Yuan, Zhang Zhang, Liang Wang, Rong Jin, and Tieniu Tan. AdaNPC: Exploring non-parametric classifier for test-time adaptation. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara En- gelhardt, Sivan Sabato, and Jonathan Scarlett, edi...

  53. [61]

    bike”, “bus

    Hao Zhao, Yuejiang Liu, Alexandre Alahi, and Tao Lin. On pitfalls of test-time adaptation. In International Conference on Machine Learning (ICML), 2023. A Appendix A.1 Sample analysis In Section 4.2, experimental results indicate that CIVD contributes the most to the im- prove...

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.