pith. sign in

arxiv: 2607.00259 · v1 · pith:IZCUEXISnew · submitted 2026-06-30 · 💻 cs.CV · cs.AI

Multi-Hypothesis Test-Time Adaptation to Mitigate Underspecification

Afshar Shamsi , Xiao-Yu Guo , Hamid Alinejad-Rokny , Arash Mohammadi , Damien Teney , Ehsan Abbasnejad This is my paper

Pith reviewed 2026-07-02 19:07 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords test-time adaptationunderspecificationdistribution shiftentropy minimizationmulti-hypothesis inferencerobustnesscomputer vision
0
0 comments X

The pith

Treating test-time adaptation as inference over multiple low-entropy hypotheses, instead of a single parameter update, reduces underspecification and yields more stable performance under distribution shifts.

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

Test-time adaptation improves a pretrained model on unlabeled target data by minimizing the entropy of its predictions. Without labels, however, many different parameter changes can reach similarly low entropy while producing very different decision boundaries. The paper argues that this underconstrained nature is the root cause of brittle behavior in standard TTA methods. It proposes exploring several such low-entropy solutions in parallel through diversification at output, parameter, optimizer, and input levels, then aggregating them. A reader would care because the approach is presented as a plug-in wrapper that improves robustness on existing benchmarks without changing the underlying adaptation objective.

Core claim

Entropy minimization during test-time adaptation defines a pseudo-likelihood over parameters, but this likelihood is underconstrained: multiple distinct parameter vectors achieve comparable low entropy yet induce different boundaries. The paper therefore reframes TTA as posterior inference over these solutions and replaces single-point optimization with a particle-based diversification procedure that simultaneously tracks multiple adaptation trajectories at four levels, producing an aggregated predictor that is less prone to collapse into spurious modes.

What carries the argument

Particle-based multi-level diversification framework that maintains and aggregates multiple plausible adaptation trajectories.

If this is right

  • Gains of 3-4% on mixed distribution shifts, 2-3% at batch size one, and 1-2.5% under label shifts.
  • The wrapper can be attached to any existing entropy-based TTA method.
  • Diversification at output, parameter, optimizer, and input levels together produce the reported stability.
  • The method treats low-entropy solutions as defining a pseudo-posterior rather than committing to one point estimate.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same multi-hypothesis view could be applied to other unsupervised objectives that suffer from multiple equally low-loss solutions.
  • Aggregation of particles may reduce sensitivity to the exact choice of entropy-minimization hyperparameters.
  • If the diversification levels interact, ablating any one level should measurably reduce the observed gains.

Load-bearing premise

That several meaningfully different low-entropy parameter updates exist and can be combined without creating new failure modes that cancel the reported gains.

What would settle it

A controlled experiment on a benchmark where every low-entropy solution found by the base TTA method produces identical predictions on the test set, or where running the multi-hypothesis wrapper yields no improvement or a measurable drop.

Figures

Figures reproduced from arXiv: 2607.00259 by Afshar Shamsi, Arash Mohammadi, Damien Teney, Ehsan Abbasnejad, Hamid Alinejad-Rokny, Xiao-Yu Guo.

Figure 1
Figure 1. Figure 1: Conceptual illustration of multi-hypothesis test-time adaptation. Entropy mini￾mization can yield multiple low-entropy solutions in the adaptation landscape. Standard TTA follows a single trajectory from θ 0 , which may converge to a suboptimal decision boundary. Our framework instead maintains multiple adaptation particles (θ1, θ2) by adapting separate normalization parameters. Aggregating their predictio… view at source ↗
Figure 2
Figure 2. Figure 2: An illustration of the proposed method for the case of gradient diversity diversification. During the adaptation, we only update the normalization layers (N) and keep the rest of layers (W and C) frozen. When a batch of test samples come, we first identify non-harmful ones (see Appendix D). We only perform backward propagation on selected samples with gradient diversity measure to push normalization layers… view at source ↗
Figure 3
Figure 3. Figure 3: The interpolation of three optima in the test loss landscape is visualized, where blue regions indicate high error/loss and red regions correspond to low error/loss. 2 Preliminaries 2.1 Test Time Adaptation Suppose that we have a training (source) dataset Ds = {(x s j , ys j )} Ns j=1 where x s j ∈ X s and y s j ∈ Ys , Ns is the number of training instances, and a testing (target) set Dt = {(x t j , yt j )… view at source ↗
Figure 4
Figure 4. Figure 4: Ablation studies of diversification design choices on ImageNet-C under the wild test-time adaptation scenario (batch size 1) using a ViTBase-LN backbone. minima corresponding to different augmented perspectives of the target stream. As reported in Tab. 5, combining input diversification with gradient-based repulsion yields the strongest robustness under batch-size-one adaptation. Across corruption types, t… view at source ↗
read the original abstract

Test-Time Adaptation (TTA) seeks to improve model robustness under distribution shifts by adapting parameters using unlabeled target data. However, in the absence of supervision, entropy-based adaptation is fundamentally underconstrained: multiple distinct parameter updates can achieve similarly low entropy while inducing drastically different decision boundaries. This phenomenon, known as underspecification, renders standard TTA brittle and prone to collapse into spurious modes. In this work, we reinterpret TTA through a posterior-inspired lens induced by entropy minimization, where low-entropy solutions define a pseudo-likelihood over parameters. Instead of committing to a single point estimate, we introduce a particle-based diversification framework that explores multiple plausible adaptation trajectories simultaneously. Our method can be viewed as a structured exploration of multiple plausible adaptation solutions, implemented through multi-level diversification at the output, parameter, optimizer, and input levels. Crucially, the framework acts as a plug-and-play wrapper compatible with existing TTA methods. Extensive experiments on challenging benchmarks demonstrate consistent gains in stability and robustness, achieving improvements of 3-4% under mixed shifts, 2-3% with batch size one, and 1-2.5% under label shifts, outperforming state-of-the-art baselines. Our results suggest that treating TTA as a multi-hypothesis inference problem, rather than a single-point optimization task, is key to mitigating underspecification and enabling reliable real-world deployment.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper claims that entropy-minimization TTA is fundamentally underspecified because multiple distinct low-entropy parameter vectors can induce different decision boundaries on unlabeled target data. It reinterprets TTA via a pseudo-likelihood over parameters and introduces a particle-based diversification framework that performs multi-level exploration (output, parameter, optimizer, input) as a plug-and-play wrapper around existing TTA methods, reporting 3-4% gains under mixed shifts, 2-3% at batch size 1, and 1-2.5% under label shifts on standard benchmarks.

Significance. If the reported gains are shown to arise specifically from capturing distinct low-entropy modes rather than ancillary ensembling effects, the multi-hypothesis framing could meaningfully improve robustness of TTA under real-world shifts. The plug-and-play design is a practical strength that would facilitate adoption if the core premise is validated.

major comments (3)
  1. [Abstract] Abstract: the claim that 'multiple distinct parameter updates can achieve similarly low entropy while inducing drastically different decision boundaries' is load-bearing for the entire multi-hypothesis argument, yet the manuscript provides no direct measurements (parameter cosine similarity, prediction disagreement rates on target batches, or boundary divergence metrics) to confirm that the particles explore meaningfully distinct modes rather than correlated solutions.
  2. [Experiments] Experiments section: the 3-4% gains under mixed shifts are reported without ablations that isolate the contribution of the multi-hypothesis aggregation step versus the multi-level diversification components alone; this leaves open whether the pseudo-likelihood reinterpretation adds explanatory power beyond standard diversification or implicit ensembling.
  3. [Method] Method section: the particle-based framework is presented as exploring 'multiple plausible adaptation trajectories,' but without quantitative verification that the particles remain in distinct low-entropy basins (e.g., via entropy histograms or mode-separation statistics across runs), the central premise that aggregation mitigates underspecification remains unverified.
minor comments (2)
  1. [Method] The phrase 'structured exploration of multiple plausible adaptation solutions' is repeated without a concise formal definition; a short paragraph or pseudocode box early in the method would improve clarity.
  2. [Tables] Table captions should explicitly state whether reported numbers are means over multiple random seeds and whether error bars or standard deviations are shown.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback highlighting the need for stronger empirical validation of the multi-hypothesis premise. We address each major comment below and will incorporate the requested analyses and ablations in the revised manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that 'multiple distinct parameter updates can achieve similarly low entropy while inducing drastically different decision boundaries' is load-bearing for the entire multi-hypothesis argument, yet the manuscript provides no direct measurements (parameter cosine similarity, prediction disagreement rates on target batches, or boundary divergence metrics) to confirm that the particles explore meaningfully distinct modes rather than correlated solutions.

    Authors: We agree that direct measurements are needed to substantiate the claim of distinct modes. In the revision we will add quantitative analyses including parameter cosine similarity between particles, prediction disagreement rates on held-out target batches, and boundary divergence metrics (e.g., via disagreement on synthetic boundary probes) to demonstrate that the particles occupy meaningfully different low-entropy solutions rather than correlated ones. revision: yes

  2. Referee: [Experiments] Experiments section: the 3-4% gains under mixed shifts are reported without ablations that isolate the contribution of the multi-hypothesis aggregation step versus the multi-level diversification components alone; this leaves open whether the pseudo-likelihood reinterpretation adds explanatory power beyond standard diversification or implicit ensembling.

    Authors: We acknowledge the value of isolating the aggregation step. The revised manuscript will include new ablation studies that (i) disable the multi-hypothesis aggregation while retaining all diversification components and (ii) compare against standard ensembling baselines, thereby quantifying the incremental benefit attributable to the pseudo-likelihood framing. revision: yes

  3. Referee: [Method] Method section: the particle-based framework is presented as exploring 'multiple plausible adaptation trajectories,' but without quantitative verification that the particles remain in distinct low-entropy basins (e.g., via entropy histograms or mode-separation statistics across runs), the central premise that aggregation mitigates underspecification remains unverified.

    Authors: We will add the requested verification. The revision will report entropy histograms across particles, mode-separation statistics (e.g., pairwise KL divergence of output distributions and basin occupancy counts over repeated runs), confirming that particles consistently occupy distinct low-entropy basins rather than collapsing to the same mode. revision: yes

Circularity Check

0 steps flagged

No circularity: conceptual reinterpretation without equations or self-referential reductions.

full rationale

The paper presents TTA as a multi-hypothesis problem via a pseudo-likelihood lens induced by entropy minimization, but this is framed as a conceptual shift rather than a derivation from equations. No mathematical steps, fitted parameters renamed as predictions, or load-bearing self-citations appear in the provided text. The central claim rests on empirical gains from diversification, which are externally falsifiable via benchmarks and do not reduce to the inputs by construction. The absence of any derivation chain means the work is self-contained against external evaluation.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the domain assumption that entropy minimization induces a useful pseudo-likelihood and that diversification across multiple trajectories yields net robustness gains without new underspecification.

axioms (1)
  • domain assumption Entropy minimization defines a pseudo-likelihood over parameters
    Explicitly stated in the abstract as the reinterpretation lens for TTA.
invented entities (1)
  • particle-based diversification framework no independent evidence
    purpose: To explore multiple plausible adaptation trajectories simultaneously
    New framework introduced to mitigate underspecification; no independent evidence provided in abstract.

pith-pipeline@v0.9.1-grok · 5801 in / 1066 out tokens · 23854 ms · 2026-07-02T19:07:45.495904+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

44 extracted references · 10 canonical work pages · 1 internal anchor

  1. [1]

    In: Precup, D., Teh, Y.W

    Arpit, D., Jastrzebski, S., Ballas, N., Krueger, D., Bengio, E., Kanwal, M.S., Ma- haraj, T., Fischer, A., Courville, A.C., Bengio, Y., Lacoste-Julien, S.: A closer look at memorization in deep networks. In: Precup, D., Teh, Y.W. (eds.) Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 201...

    2017

  2. [2]

    In: NeurIPS 2021 Competitions and Demonstrations Track

    Bashkirova, D., Hendrycks, D., Kim, D., Liao, H., Mishra, S., Rajagopalan, C., Saenko, K., Saito, K., Tayyab, B.U., Teterwak, P., et al.: Visda-2021 competition: Universal domain adaptation to improve performance on out-of-distribution data. In: NeurIPS 2021 Competitions and Demonstrations Track. pp. 66–79. PMLR (2022)

    2021

  3. [3]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Chen, D., Wang, D., Darrell, T., Ebrahimi, S.: Contrastive test-time adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 295–305 (2022)

    2022

  4. [4]

    In: 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. O...

    2021

  5. [5]

    In: Bach, F.R., Blei, D.M

    Ganin, Y., Lempitsky, V.S.: Unsupervised domain adaptation by backpropagation. In: Bach, F.R., Blei, D.M. (eds.) Proceedings of the 32nd International Conference on Machine Learning, ICML 2015, Lille, France, 6-11 July 2015. JMLR Workshop and Conference Proceedings, vol. 37, pp. 1180–1189. JMLR.org (2015),http: //proceedings.mlr.press/v37/ganin15.html

    2015

  6. [6]

    Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., Marchand, M., Lempitsky, V.S.: Domain-adversarial training of neural networks. J. Mach. Learn. Res.17, 59:1–59:35 (2016),https://jmlr.org/papers/v17/15- 239.html

    2016

  7. [7]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8340–8349 (2021)

    2021

  8. [8]

    Benchmarking Neural Network Robustness to Common Corruptions and Perturbations

    Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261 (2019)

    work page internal anchor Pith review Pith/arXiv arXiv 1903

  9. [9]

    In: 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019

    Hendrycks, D., Dietterich, T.G.: Benchmarking neural network robustness to com- mon corruptions and perturbations. In: 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenRe- view.net (2019),https://openreview.net/forum?id=HJz6tiCqYm

    2019

  10. [10]

    Shamsi et al

    Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (Jul 2021).https://doi.org/10.5281/zenodo.5143773 , 16 A. Shamsi et al. https://doi.org/10.5281/zenodo.5143773, if you use this software, please cite it as below

    work page doi:10.5281/zenodo.5143773 2021

  11. [11]

    In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=9w3iw8wDuE

    Lee, J., Jung, D., Lee, S., Park, J., Shin, J., Hwang, U., Yoon, S.: Entropy is not enough for test-time adaptation: From the perspective of disentangled factors. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=9w3iw8wDuE

    2024

  12. [12]

    arXiv preprint arXiv:2403.07366 (2024)

    Lee, J., Jung, D., Lee, S., Park, J., Shin, J., Hwang, U., Yoon, S.: Entropy is not enough for test-time adaptation: From the perspective of disentangled factors. arXiv preprint arXiv:2403.07366 (2024)

    work page arXiv 2024

  13. [13]

    CoRRabs/2303.15361(2023)

    Liang, J., He, R., Tan, T.: A comprehensive survey on test-time adaptation under distribution shifts. CoRRabs/2303.15361(2023). https://doi.org/10.48550/ ARXIV.2303.15361,https://doi.org/10.48550/arXiv.2303.15361

    work page doi:10.48550/arxiv.2303.15361 2023

  14. [14]

    In: Advances in Neural Information Processing Systems

    Liu, Q., Wang, D.: Stein variational gradient descent: A general purpose bayesian inference algorithm. In: Advances in Neural Information Processing Systems. vol. 29 (2016)

    2016

  15. [15]

    Advances in neural information processing systems29(2016)

    Liu, Q., Wang, D.: Stein variational gradient descent: A general purpose bayesian inference algorithm. Advances in neural information processing systems29(2016)

    2016

  16. [16]

    In: Proceedings of the 30th International Conference on Machine Learning, ICML 2013, Atlanta, GA, USA, 16-21 June 2013

    Muandet, K., Balduzzi, D., Schölkopf, B.: Domain generalization via invariant feature representation. In: Proceedings of the 30th International Conference on Machine Learning, ICML 2013, Atlanta, GA, USA, 16-21 June 2013. JMLR Workshop and Conference Proceedings, vol. 28, pp. 10–18. JMLR.org (2013), http://proceedings.mlr.press/v28/muandet13.html

    2013

  17. [17]

    CoRRabs/2006.10963(2020),https://arxiv.org/abs/2006.10963

    Nado, Z., Padhy, S., Sculley, D., D’Amour, A., Lakshminarayanan, B., Snoek, J.: Evaluating prediction-time batch normalization for robustness under covariate shift. CoRRabs/2006.10963(2020),https://arxiv.org/abs/2006.10963

    work page arXiv 2006

  18. [18]

    In: Wallach, H.M., Larochelle, H., Beygelzimer, A., d’Alché-Buc, F., Fox, E.B., Garnett, R

    Nagarajan, V., Kolter, J.Z.: Uniform convergence may be unable to explain gen- eralization in deep learning. In: Wallach, H.M., Larochelle, H., Beygelzimer, A., d’Alché-Buc, F., Fox, E.B., Garnett, R. (eds.) Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2...

    2019

  19. [19]

    In: Internetional Conference on Learning Representations (2023)

    Niu, S., Wu, J., Zhang, Y., Wen, Z., Chen, Y., Zhao, P., Tan, M.: Towards stable test- time adaptation in dynamic wild world. In: Internetional Conference on Learning Representations (2023)

    2023

  20. [20]

    arXiv preprint arXiv:2406.13875 (2024)

    Osowiechi, D., Noori, M., Hakim, G.A.V., Yazdanpanah, M., Bahri, A., Cher- aghalikhani, M., Dastani, S., Beizaee, F., Ayed, I.B., Desrosiers, C.: Watt: Weight average test-time adaptation of clip. arXiv preprint arXiv:2406.13875 (2024)

    work page arXiv 2024

  21. [21]

    In: International Conference on Machine Learning

    Press, O., Shwartz-Ziv, R., LeCun, Y., Bethge, M.: The entropy enigma: Success and failure of entropy minimization. In: International Conference on Machine Learning. PMLR (2024)

    2024

  22. [22]

    In: Zhou, Z

    Qiu, Z., Zhang, Y., Lin, H., Niu, S., Liu, Y., Du, Q., Tan, M.: Source-free domain adaptation via avatar prototype generation and adaptation. In: Zhou, Z. (ed.) Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI 2021, Virtual Event / Montreal, Canada, 19-27 August 2021. pp. 2921–2927. ijcai.org (2021). https://do...

    work page doi:10.24963/ijcai.2021/402 2021

  23. [23]

    MIT Press, Cambridge, MA (2009) Multi-Hypothesis TTA 17

    Quiñonero, J., Sugiyama, M., Schwaighofer, A., Lawrence, N.D.: Dataset Shift in Machine Learning. MIT Press, Cambridge, MA (2009) Multi-Hypothesis TTA 17

    2009

  24. [24]

    In: Meila, M., Zhang, T

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual E...

    2021

  25. [25]

    In: Proceedings of the 38th International Conference on Machine Learning (ICML) (2021)

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Proceedings of the 38th International Conference on Machine Learning (ICML) (2021)

    2021

  26. [26]

    Recht, B., Roelofs, R., Schmidt, L., Shankar, V.: Do imagenet classifiers generalize to imagenet? In: International conference on machine learning. pp. 5389–5400. PMLR (2019)

    2019

  27. [27]

    In: Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event

    Sagawa, S., Raghunathan, A., Koh, P.W., Liang, P.: An investigation of why overparameterization exacerbates spurious correlations. In: Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event. Proceedings of Machine Learning Research, vol. 119, pp. 8346–8356. PMLR (2020),http://proceedings.mlr.press/v...

    2020

  28. [28]

    In: 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018

    Saito, K., Watanabe, K., Ushiku, Y., Harada, T.: Maximum classifier discrepancy for unsupervised domain adaptation. In: 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018. pp. 3723–3732. Computer Vision Foundation / IEEE Computer Society (2018). https://doi.org/10.1109/CVPR.2018.00392 , ht...

    work page doi:10.1109/cvpr.2018.00392 2018

  29. [29]

    In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H

    Shah, H., Tamuly, K., Raghunathan, A., Jain, P., Netrapalli, P.: The pitfalls of simplicity bias in neural networks. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual (2020...

    2020

  30. [30]

    The Bell System Tech- nical Journal27, 379–423 (1948), http://plan9.bell-labs.com/cm/ms/what/ shannonday/shannon1948.pdf

    Shannon, C.E.: A mathematical theory of communication. The Bell System Tech- nical Journal27, 379–423 (1948), http://plan9.bell-labs.com/cm/ms/what/ shannonday/shannon1948.pdf

    1948

  31. [31]

    In: International conference on machine learning

    Sun, Y., Wang, X., Liu, Z., Miller, J., Efros, A., Hardt, M.: Test-time training with self-supervision for generalization under distribution shifts. In: International conference on machine learning. pp. 9229–9248. PMLR (2020)

    2020

  32. [32]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Teney, D., Abbasnejad, E., Lucey, S., van den Hengel, A.: Evading the simplicity bias: TrainingadiversesetofmodelsdiscoverssolutionswithsuperiorOODgeneralization. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. pp. 16740–16751. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.01626,...

    work page doi:10.1109/cvpr52688.2022.01626 2022

  33. [33]

    In: International Conference on Learning Representations (2021),https://openreview.net/forum?id=uXl3bZLkr3c

    Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. In: International Conference on Learning Representations (2021),https://openreview.net/forum?id=uXl3bZLkr3c

    2021

  34. [34]

    Advances in neural information processing systems35, 38629–38642 (2022) 18 A

    Zhang, M., Levine, S., Finn, C.: Memo: Test time robustness via adaptation and augmentation. Advances in neural information processing systems35, 38629–38642 (2022) 18 A. Shamsi et al. Multi-Hypothesis Test-Time Adaptation to Mitigate Underspecification Overview of Materials in the Appendices A brief overview of additional experimental results and finding...

    2022

  35. [35]

    Related work (Appendix A)

  36. [36]

    Underspecification in TTA (Appendix B)

  37. [37]

    Beyond Corruption-Based Shifts (Appendix C)

  38. [38]

    Sample Selection and Hyperparameter Configuration (Appendix D)

  39. [39]

    Optimization Objective of SVGD (Appendix E)

  40. [40]

    Additional experiments regarding mild scenarios (Appendix F)

  41. [41]

    Additional experiments associated with wild scenarios for severity level of 3 (Appendix G)

  42. [42]

    Runtime comparison of methods (Appendix H)

  43. [43]

    Existing research in OOD generalization primarily focuses on domain adaptation and invariant representation learning

    Limitation (Appendix I) A Related Work A.1 Out-of-distribution Generalization Out-of-distribution (OOD) generalization has become a critical research area, as models deployed in real-world scenarios often encounter data that are different from the training distribution. Existing research in OOD generalization primarily focuses on domain adaptation and inv...

  44. [44]

    [32] trained a collection of models andidentifiedonlyoneforinference,whichdiscoveredpredictivepatternsnormally missed by a learning algorithm because of the simplicity bias

    proposed techniques based on adversarial training, where models are trained against perturbed data points, encouraging them to move beyond simpler features and to learn more generalizable representations. [32] trained a collection of models andidentifiedonlyoneforinference,whichdiscoveredpredictivepatternsnormally missed by a learning algorithm because of...

    work page arXiv 2021