Pith. sign in

REVIEW 3 major objections 4 minor 49 references

When Small Guides Large: Cross-Model Co-Learning for Test-Time Adaptation

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that two pre-trained models can mutually improve test-time adaptation by co-learning, so that a much smaller model reliably guides a larger one: with Mobile-ViT as the auxiliary, ViT-Base's average ImageNet-C accuracy…

desk verdict Bidirectional co-learning for TTA is real and mostly works; add the independent-ensemble control and fix the abstract before it's final. read the letter →

arxiv 2506.23724 v2 pith:4C3E7V2T submitted 2025-06-30 cs.CV cs.AI

Chang'an Yi , Xiaohui Deng , Guohao Chen , Yan Zhou , Qinghua Lu , Shuaicheng Niu This is my paper
classification cs.CVcs.AI
keywords test-timeadaptationco-learningknowledgedistillationentropyminimizationmodelensemblesdomainshiftplug-and-playtemperaturescaling
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 asks whether two pre-trained models can help each other adapt to out-of-distribution test data in an online, unsupervised setting, and claims the answer is yes, even when the models differ enormously in size. It introduces COCA, which lets each model both learn from the other and adapt on its own, so a 10.6M-parameter Mobile-ViT can act as a reliable guide for an 86.6M-parameter ViT-Base. Across ImageNet-C and real-world shifts, COCA consistently beats single-model test-time adaptation and can be bolted onto existing methods as a plug-and-play module. The headline result is an increase in ViT-Base's average ImageNet-C accuracy from 51.7% with a standard entropy-minimization baseline to 64.5% with COCA.

What carries the argument

The load-bearing device is a learnable, diversity-aware temperature τ, updated to minimize the L1 discrepancy between the anchor's exponential logits $e^{p_a(x)}$ and the auxiliary's scaled exponential logits $e^{p_s(x)/\tau}$. This τ is then used to aggregate predictions as $p_e = (p_a + p_s/\tau)/T$, where T keeps the max logit unchanged. Around this ensemble, COCA builds three losses: marginal entropy minimization $\mathcal{L}_{mar}$ on $p_e$, cross-model knowledge distillation $\mathcal{L}_{ckd}$ using $p_e$'s argmax as pseudo-label for both models, and self-adaptation $\mathcal{L}_{sa}$ via per-model entropy minimization, with total loss $\mathcal{L} = \mathcal{L}_{col} + \mathcal{L}_{sa}$. The mechanism is what lets a smaller, more noise-robust model influence a larger, more accurate but less robust one without the larger model being dragged down.

What would settle it

Run COCA on a corruption where the anchor is confidently wrong and the auxiliary is correct but less confident, or on a synthetic shift where both models output the same wrong class; if the L1 discrepancy fails to raise τ and the combined accuracy drops below the better single model, the trust signal is not reliable.

Watch

Extended reading notes

Core claim

The central claim is that complementary and robustness-differential knowledge between two pre-trained models can be harvested during test-time adaptation, bidirectionally, rather than only in a one-way teacher-student transfer. COCA designates the larger model as the anchor and the other as the auxiliary, and learns a scaling temperature τ that aligns the auxiliary's exponential logits with the anchor's before forming the ensemble prediction $p_e = (p_a + p_s/\tau)/T$. That ensemble drives a marginal entropy minimization loss, while a cross-model knowledge distillation loss uses the ensemble's pseudo-label to supervise both models, and each model also keeps its own entropy-based self-adaptation. The paper reports that with Mobile-ViT as auxiliary, ViT-Base reaches 64.0% and the combined prediction 64.5%, and that COCA lifts a sample-filtered TTA baseline from 62.0% to 69.1% as a plug-in.

Load-bearing premise

The method assumes that the L1 difference between the anchor's exponential logits and the auxiliary's scaled exponential logits reliably measures how much the auxiliary should be trusted in the ensemble; if that signal misleads under severe or correlated domain shift, τ would scale the wrong model and the gain would disappear.

Editorial extensions

If this is right

  • With ResNet-50 as auxiliary, combined COCA reaches 64.9% on ImageNet-C, while each individual model also improves over its single-model adaptation.
  • As a plug-and-play module, COCA raises the sample-filtered EATA baseline from 62.0% to 69.1% and improves SAR-based adaptation, with only modest GPU time and memory overhead.
  • If one model collapses under label shift, τ grows and suppresses that model's contribution, so the combined performance stays near the surviving model.
  • COCA extends to three models by hierarchically treating pairs as anchor and auxiliary, and the extra small model usually boosts the two larger ones.
  • Across architecture pairs, including Mobile-ViT with ViT-Large and same-architecture ViTs with different weights, both models consistently beat their single-model baselines.

Reading between the lines

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

  • A natural stress test is to track how much prediction diversity actually survives mutual distillation; if the two models converge to identical outputs, the co-learning gain should vanish, and the paper does not directly measure this.
  • Because τ is learned from prediction disagreement in exponential space, it could plausibly serve as an online confidence gate for other auxiliary signals, such as augmentations or test-time prompts.
  • A theoretical account of when the L1 discrepancy is a valid estimator of auxiliary trustworthiness would turn this heuristic into a principled component; until then, its failure modes under severe correlated shifts remain open.
  • The plug-and-play gains suggest that even a single deployed model could benefit from a small, cheap companion that only helps when the main model's confidence is low.
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

3 major / 4 minor

Summary. The paper proposes COCA, a test-time adaptation (TTA) framework that adapts two pre-trained models jointly. A learnable temperature τ scales the auxiliary model's logits before they are added to the anchor model's logits, and the combined prediction is used for entropy minimization and knowledge-distillation losses, while each model also applies its own entropy minimization. Experiments on ImageNet-C, OfficeHome, ImageNet-R/Sketch, and CIFAR-100-C report consistent improvements over single-model TTA baselines and show that a smaller model such as Mobile-ViT can improve the adaptation accuracy of a larger model such as ViT-Base (from 51.7% to 64.0% for ViT-Base itself, and to 64.5% for the combined ensemble).

Significance. If the claims hold, this is a useful contribution: a simple, computationally light mechanism for exploiting complementarity between two models during online unsupervised adaptation, with promising plug-and-play behavior on existing entropy-based TTA methods. The evaluation is broad—15 corruption types, several datasets, all pairwise combinations of six models—and the code is promised publicly. However, the central scientific claim is about mutual co-adaptation, not merely about having two models, and the experiments do not yet separate these two factors. The paper also does not provide statistical reliability for its numbers, and the central τ mechanism remains a heuristic. These gaps prevent me from endorsing the results as they stand, despite the extensive empirical effort.

major comments (3)
  1. [§4.1, Fig. 3, Table 12]
  2. [§3.1, Eq. (2)–(3)]
  3. [§4.1, Tables 1 and 12]
minor comments (4)
  1. [Abstract]
  2. [Appendix C, Table 11]
  3. [§4.2, Fig. 4]
  4. [§1, Introduction]

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: COCA's losses and temperature parameter are optimized on unlabeled test batches and evaluated on held-out labels; no fitted quantity is renamed as a prediction.

full rationale

The paper's derivation chain is self-contained in the relevant sense: the co-adaptation and self-adaptation losses (Eqs. 4-8) are optimized online on unlabeled target data, and the reported accuracy numbers are measured against held-out corruption labels. The learnable temperature τ is fit to an unsupervised L1 discrepancy loss (Eq. 2) and to the resulting ensemble logits (Eq. 3); it is not fit to labels, nor is any label-derived quantity used to define the training objective. The reported gains are therefore empirical outcomes, not consequences of a fitted parameter being renamed as a prediction. The included 'Average' baseline in Fig. 3, which jointly minimizes entropy of averaged logits without τ, partially controls for the effect of the adaptive scaling; although a fully independent two-model ensemble control is not reported, that is an experimental-completeness concern, not a circularity. The authors do cite their own prior work (EATA [28], SAR [29]) and use those methods as baselines and as plug-and-play backbones, but the core COCA mechanism does not depend on an unverified self-citation; Tent's entropy minimization, which is external, supplies the base objective. One reporting inaccuracy exists: the abstract states that "COCA raises ViT-Base's average adaptation accuracy on ImageNet-C from 51.7% to 64.5%," whereas Table 12 lists ViT-Base* at 64.0% and the 'Combined' ensemble at 64.5%; this conflates ensemble accuracy with the anchor model's individual accuracy, but as a numbers-reporting issue it is not a circularity of the derivation. Overall, no equation reduces to its own input and no fitted parameter is presented as a held-out prediction.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

No new physical entities are introduced. The method relies on standard TTA assumptions and one ad hoc scaling heuristic (tau driven by L1 discrepancy). The free parameters are all online-learned or low-sensitivity hyperparameters.

free parameters (3)
  • temperature scale tau (per-batch learnable) = initialized 1, updated by Eq. (1) to minimize L1 discrepancy; ranges roughly 1-5 in experiments
    Controls how much the auxiliary model contributes to the ensemble; learned online, not fitted to labels. It is a per-batch scalar that adapts over time, so it is a learned parameter, not a fixed constant.
  • K (update steps for tau) = default 5 (3-5 suffice)
    Hyperparameter for inner-loop updates of tau; authors show low sensitivity across K.
  • loss ratio L_col : L_sa = 1:1
    Weighting of co-adaptation vs. self-adaptation losses; authors note 1:2 gives slight gain (69.1 to 69.5 on EATA+COCA) but choose 1:1 for simplicity.
assumptions (5)
  • domain assumption Both models are pre-trained on the same source distribution (ImageNet) and share the same label space.
    Problem statement in Section 3 defines both f_theta1 and f_theta2 as well-trained on Dtrain; this makes complementary knowledge meaningful. If models came from different source domains, co-learning behavior could differ.
  • domain assumption Entropy minimization provides a valid unsupervised objective for TTA.
    COCA builds on Tent's entropy minimization; the validity of the objective is taken from prior work [36].
  • ad hoc to paper The L1 discrepancy in exponential space between anchor and auxiliary predictions is a reliable signal for scaling the auxiliary contribution.
    Eq. (2) defines this discrepancy without theoretical justification; the success of tau depends on this heuristic. This is the weakest assumption.
  • domain assumption Only batch normalization layers are adapted, following Tent.
    Implementation details state this setup; limits adaptation capacity but follows the baseline.
  • ad hoc to paper The adaptive balance factor T preserves the anchor's max logit, maintaining a reasonable sharpness of p_e.
    T is defined as max p'_e / max p_a and used in Eq. (3); it is a heuristic to avoid softening the ensemble.

how reviews work

0 comments
Cite this review

Pith. "Pith review of When Small Guides Large: Cross-Model Co-Learning for Test-Time Adaptation." pith.science (2026). https://pith.science/paper/4C3E7V2T

@misc{pith2026250623724,
  author       = {Pith},
  title        = {Pith review of: When Small Guides Large: Cross-Model Co-Learning for Test-Time Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4C3E7V2T}},
  note         = {Machine review of arXiv:2506.23724}
}
read the original abstract

Test-time Adaptation (TTA) adapts a given model to testing domain data with potential domain shifts through online unsupervised learning, yielding impressive performance. However, to date, existing TTA methods primarily focus on single-model adaptation. In this work, we investigate an intriguing question: how does cross-model knowledge influence the TTA process? Our findings reveal that, in TTA's unsupervised online setting, each model can provide complementary, confident knowledge to the others, even when there are substantial differences in model size. For instance, a smaller model like MobileViT (10.6M parameters) can effectively guide a larger model like ViT-Base (86.6M parameters). In light of this, we propose COCA, a Cross-Model Co-Learning framework for TTA, which mainly consists of two main strategies. 1) Co-adaptation adaptively integrates complementary knowledge from other models throughout the TTA process, reducing individual model biases. 2) Self-adaptation enhances each model's unique strengths via unsupervised learning, enabling diverse adaptation to the target domain. Extensive experiments show that COCA, which can also serve as a plug-and-play module, significantly boosts existing SOTAs, on models with various sizes--including ResNets, ViTs, and Mobile-ViTs--via cross-model co-learned TTA. For example, with Mobile-ViT's guidance, COCA raises ViT-Base's average adaptation accuracy on ImageNet-C from 51.7% to 64.5%. The code is publicly available at https://github.com/ycarobot/COCA.

Figures

Figures reproduced from arXiv: 2506.23724 by the authors.

Figure 1
Figure 1. Motivation for COCA. (a) Pre-trained models exhibit distinct strengths due to differences in training strategies, architectures, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The overall COCA framework consists of two mod [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The necessity of introducing τ , a learnable parameter. All experiments are based on Tent [36]. Individual refers to adapting each model independently, as done in Tent. Average involves combining the predictions of two models by averaging their output logits for marginal entropy minimization. Under this strategy, the performance improvement is limited. In contrast, Adaptively scaled utilizes the parameter τ to adapt… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Robustness of COCA stems from τ . In the label-shift scenario [29], COCA maintains high performance even when one of the models—such as ResNet-50 in this figure—collapses. The evaluation is conducted on ImageNet-C with Gaussian noise (%). Models Accuracy (%) Parameters…
Figure 5
Figure 5. Figure 5: A sample-level analysis highlights the advantages of our proposed cross-model co-learning approach. For each sample, we select [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 38 canonical work pages

  1. [1]

    Knowledge distilla- tion: A good teacher is patient and consistent

    Lucas Beyer, Xiaohua Zhai, Am´elie Royer, Larisa Markeeva, Rohan Anil, and Alexander Kolesnikov. Knowledge distilla- tion: A good teacher is patient and consistent. In Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10925–10934, 2022. 3

    work page 2022

  2. [2]

    Contrastive mean teacher for domain adaptive ob- ject detectors

    Shengcao Cao, Dhiraj Joshi, Liang-Yan Gui, and Yu-Xiong Wang. Contrastive mean teacher for domain adaptive ob- ject detectors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 23839– 23848, 2023. 3

    work page 2023

  3. [3]

    Contrastive test-time adaptation

    Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 295–305, 2022. 2

    2022

  4. [4]

    Harmo- nious teacher for cross-domain object detection

    Jinhong Deng, Dongli Xu, Wen Li, and Lixin Duan. Harmo- nious teacher for cross-domain object detection. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23829–23838, 2023. 2, 3

    work page 2023

  5. [5]

    Cluster alignment with a teacher for unsupervised domain adaptation

    Zhijie Deng, Yucen Luo, and Jun Zhu. Cluster alignment with a teacher for unsupervised domain adaptation. In Pro- ceedings of the IEEE/CVF international conference on com- puter vision, pages 9944–9953, 2019. 2

    work page 2019

  6. [6]

    Robust mean teacher for continual and gradual test-time adaptation

    Mario D ¨obler, Robert A Marsden, and Bin Yang. Robust mean teacher for continual and gradual test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7704–7714, 2023. 2, 3

    work page 2023

  7. [7]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2

    arXiv 2010

  8. [8]

    Decorate the newcomers: Visual domain prompt for continual test time adaptation

    Yulu Gan, Yan Bai, Yihang Lou, Xianzheng Ma, Renrui Zhang, Nian Shi, and Lin Luo. Decorate the newcomers: Visual domain prompt for continual test time adaptation. In Proceedings of the AAAI Conference on Artificial Intel- ligence, pages 7595–7603, 2023. 3

    work page 2023

Show all 49 references
  1. [9]

    Unsupervised domain adaptation via deep conditional adap- tation network

    Pengfei Ge, Chuan-Xian Ren, Xiao-Lin Xu, and Hong Yan. Unsupervised domain adaptation via deep conditional adap- tation network. Pattern Recognition, 134:109088, 2023. 1

    2023

  2. [10]

    Semi-supervised learning by entropy minimization

    Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. In Advances in neural information processing systems, 2004. 4

    2004

  3. [11]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 2, 5

    2016

  4. [12]

    Benchmarking neu- ral network robustness to common corruptions and perturba- tions

    Dan Hendrycks and Thomas Dietterich. Benchmarking neu- ral network robustness to common corruptions and perturba- tions. In International Conference On Learning Representa- tions, pages 1–11, 2019. 5

    2019

  5. [13]

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

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

    2021

  6. [14]

    Rethinking spa- tial dimensions of vision transformers

    Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spa- tial dimensions of vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11936–11945, 2021. 3

    2021

  7. [15]

    Distill- ing the knowledge in a neural network

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distill- ing the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 1

    2015 arXiv

  8. [16]

    Teacher-student architecture for knowledge learn- ing: A survey

    Chengming Hu, Xuan Li, Dan Liu, Xi Chen, Ju Wang, and Xue Liu. Teacher-student architecture for knowledge learn- ing: A survey. arXiv preprint arXiv:2210.17332, 2022. 2, 3

    2022 arXiv

  9. [17]

    Reliable test-time adaptation via agreement-on-the- line

    Eungyeup Kim, Mingjie Sun, Aditi Raghunathan, and Zico Kolter. Reliable test-time adaptation via agreement-on-the- line. arXiv preprint arXiv:2310.04941, 2023. 4

    2023 arXiv

  10. [18]

    Entropy is not enough for test-time adaptation: From the perspective of disentangled factors

    J Lee. Entropy is not enough for test-time adaptation: From the perspective of disentangled factors. In ICLR, 2024. 1, 3, 6, 7, 8

    2024

  11. [19]

    Entropy is not enough for test-time adaptation: From the perspec- tive of disentangled factors

    Jonghyun Lee, Dahuin Jung, Saehyung Lee, Junsung Park, Juhyeon Shin, Uiwon Hwang, and Sungroh Yoon. Entropy is not enough for test-time adaptation: From the perspec- tive of disentangled factors. In International Conference on Learning Representations, pages 1–14, 2024. 6

    2024

  12. [20]

    A comprehensive survey on source-free domain adap- tation

    Jingjing Li, Zhiqi Yu, Zhekai Du, Lei Zhu, and Heng Tao Shen. A comprehensive survey on source-free domain adap- tation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(8):5743–4762, 2024. 1

    2024

  13. [21]

    Lwsis: Lidar-guided weakly supervised instance segmentation for autonomous driving

    Xiang Li, Junbo Yin, Botian Shi, Yikang Li, Ruigang Yang, and Jianbing Shen. Lwsis: Lidar-guided weakly supervised instance segmentation for autonomous driving. In Proceed- ings of the AAAI Conference on Artificial Intelligence, pages 1433–1441, 2023. 1

    2023

  14. [22]

    A comprehensive sur- vey on test-time adaptation under distribution shifts

    Jian Liang, Ran He, and Tieniu Tan. A comprehensive sur- vey on test-time adaptation under distribution shifts. Inter- national Journal of Computer Vision, pages 1–34, 2024. 1, 2

    2024

  15. [23]

    Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems , 34: 21808–21820, 2021

    Yuejiang Liu, Parth Kothari, Bastien Van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems , 34: 21808–21820, 2021. 1 9

    2021

  16. [24]

    Understanding the robustness in vision transform- ers

    Pan Lu, Liang Qiu, Wenhao Yu, Sean Welleck, and Kai-Wei Chang. Understanding the robustness in vision transform- ers. In Annual Meeting of the Association for Computational Linguistics, pages 14605–14631, 2023. 1

    2023

  17. [25]

    Univer- sal test-time adaptation through weight ensembling, diver- sity weighting, and prior correction

    Robert A Marsden, Mario D ¨obler, and Bin Yang. Univer- sal test-time adaptation through weight ensembling, diver- sity weighting, and prior correction. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2555–2565, 2024. 3, 6, 7, 8

    2024

  18. [26]

    Im- proved knowledge distillation via teacher assistant

    Seyed Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Im- proved knowledge distillation via teacher assistant. In Pro- ceedings of the AAAI conference on artificial intelligence , pages 5191–5198, 2020. 1

    2020

  19. [27]

    Obtaining well calibrated probabilities using bayesian binning

    Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using bayesian binning. In Proceedings of the AAAI conference on artificial intelligence, 2015. 4

    2015

  20. [28]

    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 Interna- tional conference on machine learning, pages 16888–16905. PMLR, 2022. 1, 2, 3, 6, 7, 8

    2022

  21. [29]

    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. arXiv preprint arXiv:2302.12400, 2023. 1, 3, 6, 7, 8

    2023 arXiv

  22. [30]

    Test-time model adaptation with only forward passes

    Shuaicheng Niu, Chunyan Miao, Guohao Chen, Pengcheng Wu, and Peilin Zhao. Test-time model adaptation with only forward passes. arXiv preprint arXiv:2404.01650, 2024. 3

    2024 arXiv

  23. [31]

    Unsupervised intra-domain adaptation for se- mantic segmentation through self-supervision

    Fei Pan, Inkyu Shin, Francois Rameau, Seokju Lee, and In So Kweon. Unsupervised intra-domain adaptation for se- mantic segmentation through self-supervision. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3764–3773, 2020. 1

    2020

  24. [32]

    Acdc: The adverse conditions dataset with correspondences for se- mantic driving scene understanding

    Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for se- mantic driving scene understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 10765–10775, 2021. 1

    2021

  25. [33]

    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, pages 9229– 9248, 2020. 1

    2020

  26. [34]

    Post-hoc uncer- tainty calibration for domain drift scenarios

    Christian Tomani, Sebastian Gruber, Muhammed Ebrar Er- dem, Daniel Cremers, and Florian Buettner. Post-hoc uncer- tainty calibration for domain drift scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 10124–10132, 2021. 4

    2021

  27. [35]

    Deep hashing network for unsupervised domain adaptation

    Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5018–5027, 2017. 5

    2017

  28. [36]

    Tent: Fully test-time adaptation by entropy minimization

    Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Ol- shausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726,

    2006 arXiv

  29. [37]

    Learning robust global representations by penalizing local predictive power.Advances in Neural Information Pro- cessing Systems, 32, 2019

    Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power.Advances in Neural Information Pro- cessing Systems, 32, 2019. 5

    2019

  30. [38]

    Continual test-time domain adaptation

    Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7201–7211, 2022. 2, 3, 6, 8, 1

    2022

  31. [39]

    A survey on ensemble learning under the era of deep learning

    Yongquan Yang, Haijun Lv, and Ning Chen. A survey on ensemble learning under the era of deep learning. Artificial Intelligence Review, 56(6):5545–5589, 2023. 3

    2023

  32. [40]

    Crossmatch: Source- free domain adaptive semantic segmentation via cross-modal consistency training

    Yifang Yin, Wenmiao Hu, Zhenguang Liu, Guanfeng Wang, Shili Xiang, and Roger Zimmermann. Crossmatch: Source- free domain adaptive semantic segmentation via cross-modal consistency training. In Proceedings of the IEEE/CVF In- ternational Conference on Computer Vision , pages 21...

    2023

  33. [41]

    Robust test- time adaptation in dynamic scenarios

    Longhui Yuan, Binhui Xie, and Shuang Li. Robust test- time adaptation in dynamic scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15922–15932, 2023. 3

    2023

  34. [42]

    Tea: Test-time energy adaptation

    Yige Yuan, Bingbing Xu, Liang Hou, Fei Sun, Huawei Shen, and Xueqi Cheng. Tea: Test-time energy adaptation. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1–11, 2024. 2

    2024

  35. [43]

    Memo: Test time robustness via adaptation and augmentation

    Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. Ad- vances in neural information processing systems, 35:38629– 38642, 2022. 3

    2022

  36. [44]

    Deep mutual learning

    Ying Zhang, Tao Xiang, Timothy M Hospedales, and Huchuan Lu. Deep mutual learning. In Proceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 4320–4328, 2018. 1

    2018

  37. [45]

    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 International Conference on Machine Learning , pages 41647–41676, 2023. 2

    2023

  38. [46]

    Domain-specific risk minimization for domain generaliza- tion

    Yi-Fan Zhang, Jindong Wang, Jian Liang, Zhang Zhang, Baosheng Yu, Liang Wang, Dacheng Tao, and Xing Xie. Domain-specific risk minimization for domain generaliza- tion. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages 3409– 3421, 2023. 2

    2023

  39. [47]

    Domain generalization: A survey

    Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4396–4415, 2022. 1

    2022

  40. [48]

    Adaptive mutual learning for unsupervised do- main adaptation

    Lihua Zhou, Siying Xiao, Mao Ye, Xiatian Zhu, and Shuaifeng Li. Adaptive mutual learning for unsupervised do- main adaptation. IEEE Transactions on Circuits and Systems for Video Technology, 33(11):6622–6634, 2023. 2, 3 10 When Small Guides Large: Cross-Model Co-Learning for T...

    2023

  41. [2020]

    1, 2, 3, 4, 5, 6, 7, 8

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.