Pith. sign in

REVIEW 5 major objections 5 minor 43 references

Learning to Generate Gradients for Test-Time Adaptation via Test-Time Training Layers

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

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

Pith's one-line read A learned gradient generator replaces hand-designed optimizers in test-time adaptation, turning noisy unsupervised gradients into stable updates.

desk verdict A plausible L2O-for-TTA idea with real gains, but the "128 unlabeled samples" claim is weakened by supervised checkpoint selection on the same ImageNet-C validation set. read the letter →

arxiv 2412.16901 v1 pith:FPDBPLV2 submitted 2024-12-22 cs.LG cs.CV

Qi Deng , Shuaicheng Niu , Ronghao Zhang , Yaofo Chen , Runhao Zeng , Jian Chen , Xiping Hu This is my paper · ORCID
classification cs.LGcs.CV
keywords test-timeadaptationlearningtooptimizemetagradientgeneratormemorylayeronlineentropyminimizationImageNet-C
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

This paper claims that unreliable gradients during online test-time adaptation can be fixed by replacing the hand-designed optimizer with a small learned network that generates better gradients. The learned optimizer, called the Meta Gradient Generator (MGG), incorporates a gradient memory layer that compresses the history of past gradients into its parameters and then outputs a refined gradient for the current step. The authors report that pretraining MGG on just 128 unlabeled ImageNet-C validation images is enough for it to outperform existing adaptation methods on ImageNet-C, ImageNet-R, ImageNet-Sketch, and ImageNet-A, with faster convergence and lower compute budgets.

What carries the argument

The central mechanism is the Gradient Memory Layer (GML), a compact sequence model that treats every adapted model parameter's gradient history as a time series and stores that history in network parameters rather than in a hidden state vector. Each parameter has its own linear GML of hidden size 8, the layer is updated by a reconstruction loss with an adaptive learning rate, and then the frozen projection and normalization head computes the optimized gradient used to update the model's normalization layers.

What would settle it

Train MGG on a truly disjoint source of gradients, such as clean ImageNet images or only one corruption type, then deploy it on all 15 ImageNet-C corruptions; if accuracy falls to or below the SGD-based FOA baseline, the claimed cross-domain transfer of the learned optimizer is refuted. A second check is to shuffle or reverse the order of gradient batches during adaptation and observe whether the memorization claim breaks down.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that a learning-to-optimize based gradient generator can replace manually designed optimizers in test-time adaptation and, by memorizing historical gradients, turn noisy unsupervised gradients into more reliable update directions. The two-step MGG first updates its gradient memory layer through a self-supervised reconstruction loss that encodes the incoming gradient, then generates an optimized gradient through a projection, layer-normalized, tanh-activated function. The paper reports 71.3% average accuracy on ImageNet-C (severity 5), a 7.4-point gain over SAR and a 3.1-point gain over DeYO, plus higher accuracy on ImageNet-R, Sketch, and A, while using only a small pretraining set of 128 images.

Load-bearing premise

The method assumes that 128 unlabeled images from the ImageNet-C validation set are representative enough of all deployment distributions that the learned gradient corrections transfer; if the pretraining distribution does not resemble the test stream, the learned corrections could be worse than a standard optimizer.

Editorial extensions

If this is right

  • Test-time adaptation no longer needs hand-tuned optimizers or sample-filtering thresholds; the same learned generator is used across corruption and domain shifts.
  • A 5-second adaptation budget reaches 68.7% accuracy on ImageNet-C, above what competing methods achieve with their full budgets.
  • Using 35 of 782 batches (4.5% of data) already gives 69.3% accuracy, so the method is usable when test data or compute is scarce.
  • The extra GPU memory over the base adaptation loop is about 28 MB, so the learned optimizer adds little overhead.
  • The generator trained on ImageNet-C validation transfers to visually different ImageNet variants, including artistic renditions and sketches.

Reading between the lines

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

  • The paper's success suggests learned optimizers may transfer across backbones and shift types, but this is not tested; a sharper test would pretrain MGG on source-domain ImageNet images or on one corruption type and evaluate on the rest.
  • If historical-gradient memorization is the load-bearing mechanism, then shuffling or resetting gradient order during adaptation should significantly degrade MGG; the paper only compares GML against LSTM and does not report such a perturbation.
  • One testable extension is to apply GML-based gradient generation to other online learning problems, such as continual learning or video TTA, where noisy optimization signals are also a bottleneck.
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 / 5 minor

Summary. The paper proposes MGTTA, a test-time adaptation method that replaces hand-designed optimizers (e.g., SGD) with a learned optimizer called the Meta Gradient Generator (MGG). The MGG contains a Gradient Memory Layer (GML) that compresses historical gradient information into network parameters via a self-supervised reconstruction loss, and then generates refined gradients for updating the model. The MGG is pretrained on a small number (128) of samples from the ImageNet-C validation set and then deployed to unseen test streams. Experiments on ImageNet-C, ImageNet-R, ImageNet-Sketch, and ImageNet-A report accuracy improvements over existing TTA baselines, along with faster convergence and lower adaptation budgets.

Significance. If the claims hold, this is one of the first attempts to bring learning-to-optimize into online test-time adaptation, and the proposed GML is a lightweight mechanism for memorizing gradient history. The paper provides a clear algorithmic description, ablations, and efficiency analyses, and the core idea of learning to generate gradients for TTA is interesting. However, the experimental protocol raises serious concerns about whether the reported gains are actually due to the learned gradient generator or to the pretraining procedure, the choice of TTA loss, and the hyperparameter setup. The central claim that 128 unlabeled samples suffice is not fully supported, because checkpoint selection during pretraining uses labeled validation data.

major comments (5)
  1. [Appendix B.2] The pretraining protocol uses labels for model selection: 'Every 64 iterations, we randomly reinitialize the memory parameters phi_m of MGG and perform an evaluation on the validation set, ultimately selecting the MGG with the best evaluation results for TTA.' This evaluation requires ground-truth labels, which contradicts the claim that MGG is trained with 'a small number of unlabeled samples' (Abstract and Section 3.4). Because the validation set belongs to the same benchmark family as the test sets (ImageNet-C test, and closely related ImageNet variants), this selection procedure means the final MGG is not a purely unlabeled-trained model and the meta-generalization claim is not blind. Please clarify whether labels are used for selection, and if so, report the performance of the MGG without such selection.
  2. [Section 4.2 and Algorithm 1] During pretraining, Algorithm 1 updates the target model parameters theta with optimized gradients for T=2,000 iterations, and Section 4.2 states that the learning rate for theta is 1e-4 during pretraining. The paper does not state that theta is reinitialized to the original ImageNet-1K pretrained weights before TTA. If the theta used for TTA is the one that was already adapted on 128 images from the ImageNet-C validation set, then the comparison with baselines (which start from the original pretrained model) is confounded by target-domain adaptation during pretraining. Please specify explicitly whether theta is reset, and if it is, describe the exact protocol; if it is not, this is a significant leakage issue that undermines the reported improvements.
  3. [Table 4 and Section 4.4] The ablation 'Ours remove MGG' reports 70.0 accuracy on ImageNet-C, whereas the FOA baseline in Table 1, which also uses Eq. (7) as the TTA loss, reports 66.3. This 3.7-point gap is not explained. Since 'Ours remove MGG' is described as directly updating model parameters via backpropagated gradients (presumably with SGD), the comparison suggests that most of the gain over FOA might come from the optimizer/hyperparameter setup rather than from MGG itself. The paper should report the exact configuration of 'Ours remove MGG' (optimizer, learning rate, momentum, number of updates) and reconcile the difference with the FOA baseline, so readers can isolate the contribution of MGG.
  4. [All experimental tables] All tables report single runs without error bars or multiple seeds. The improvements that are central to the claims are small: the MGG ablation gain is 1.3 points (71.3 vs 70.0 in Table 4) and the gain over the strongest baseline DeYO is 3.1 points (Table 1). Without variance estimates, it is impossible to assess whether these differences are statistically significant. Please provide means and standard deviations over at least three random seeds, and ideally report the seed-dependent variability of the pretraining procedure as well.
  5. [Abstract and Section 4.3] The abstract claims a '4.2 times faster adaptation speed' compared with SAR on ImageNet-C, but this number is not directly supported by the reported runtime tables. Table 5 shows MGTTA's full runtime is 125.5s versus SAR's 242.7s (about 1.9x), and Table 2 reports accuracy under time budgets but does not define a speedup metric. Please specify how the 4.2x speedup is computed (e.g., time to reach a given accuracy threshold, or total wall-clock time) and ensure the claim is consistent with the tables.
minor comments (5)
  1. [Abstract and Section 3.4] The phrase 'unlabeled samples' is used to describe the pretraining data, but the validation set of ImageNet-C actually has labels and is used for checkpoint selection. Rephrase the claims to avoid ambiguity.
  2. [Algorithm 1] The loop indexing is confusing: line 7 computes LGML on gt-1, but the algorithm updates phi_m with the gradient of LGML with respect to phi_m. Please clarify the ordering of memorization and optimization steps, especially for t=1 where no gt-1 exists.
  3. [Eq. (7)] The entropy term sums over x in X and c in C, which yields a sum over samples and classes; this should be normalized by the batch size or number of samples to match standard implementations. Please specify the exact normalization.
  4. [Table 4 caption] The caption reads 'Effect of MGG and GML' but the table also includes an LSTM variant. Please clarify the LSTM setup (hidden size, training procedure) used in this ablation.
  5. [Section 4.3] The text says 'Compared to T3A, TENT, and CoTTA, our method achieves an average performance improvement of approximately 10% on ImageNet-C.' The actual gap from Table 1 is about 11.7, 11.7, and 9.6 points respectively, so the 'approximately 10%' should be stated more precisely.

Circularity Check

2 steps flagged · score 6.0 of 10

ImageNet-C performance is partly a product of test-set hyperparameter selection and benchmark-informed checkpoint selection, making the headline SOTA claim partially fitted rather than blind.

  1. fitted input called prediction [Appendix C.3, Table 11]
    "To choose the best learning rate for our method during TTA and to verify whether the superior performance of our method is due to the optimization ability of MGG on gradients rather than different learning rates, we conducted experiments on ImageNet-C. The results in Table 11 demonstrate that with optimal learning rates, our method outperforms Ours w/o MGG by 1.3% (71.3% vs. 70.0%)."

    The learning rate used in the reported ImageNet-C results (Table 1) is selected as the best value on the same ImageNet-C evaluation set: Table 11 sweeps learning rates on ImageNet-C and reports 71.3% at the optimal value. The headline 71.3% is therefore the argmax of a hyperparameter sweep over the very benchmark whose results are presented as the method's prediction, not an unbiased evaluation on unseen data. Comparing this selected number against baselines that use their own fixed learning rates is a fitted-input-called-prediction pattern; the margin over prior methods is partly manufactured by test-set model selection.

  2. fitted input called prediction [Appendix B.2]
    "Every 64 iterations, we randomly reinitialize the memory parameters ϕm of MGG and perform an evaluation on the validation set, ultimately selecting the MGG with the best evaluation results for TTA."

    The main text claims that 'the entire training process of MGG only requires a few number of unlabeled test samples' and that 128 unlabeled ImageNet-C validation images suffice (Section 3.4). Appendix B.2 shows the deployed MGG is not the direct product of that unlabeled training: its memory parameters are repeatedly reinitialized and the checkpoint with the best evaluation on the ImageNet-C validation set is selected. This benchmark-informed checkpoint selection means the '128 unlabeled samples' claim is not blind: the final MGG is chosen using performance on the same ImageNet-C benchmark family on which it is then evaluated, so the reported ImageNet-C results are partially forced by the selection criterion.

full rationale

The core MGG/GML derivation (Eqs. 2-7) is not circular: MGG is a learned optimizer trained with the same TTA loss it later minimizes, which is the intended learning-to-optimize scheme rather than an equation-level self-reference, and the GML reconstruction loss is self-supervised and does not encode the target result. The FOA loss is a same-author citation (Niu et al. 2024), but it is an externally published baseline and the no-MGG ablation independently shows that the loss alone accounts for much of the gain, so the self-citation is not load-bearing circularity. The circularity concern is confined to the evaluation protocol: the reported ImageNet-C accuracy is selected by sweeping the learning rate on the same ImageNet-C benchmark, and the deployed MGG checkpoint is selected by evaluation on the ImageNet-C validation set despite the paper's 'unlabeled samples only' claim. Both steps make the headline ImageNet-C numbers partially selected statistics rather than blind predictions. Generalization to ImageNet-R/Sketch/A is less directly affected because those datasets were not used for selection, though the shared test-selected hyperparameters still transfer, so the overall circularity is partial rather than total.

Assumptions & free parameters 7 free parameters · 5 assumptions · 2 invented entities

The central claim rests on a small number of learned or hand-chosen hyperparameters (hidden size, learning rate, lambda, pretraining budget) and on the domain assumption that the unsupervised TTA loss plus 128 ImageNet-C validation images provide a transferable training signal. No new physical entities are postulated; MGG and GML are new neural components with empirical evidence.

free parameters (7)
  • GML hidden size d = 8
    Selected on ImageNet-C validation (Table 10); d=2 collapses accuracy to 4.1%, d>=4 gives 71.2-71.3%.
  • TTA learning rate eta for theta = 1e-3
    Chosen via validation sweep (Table 11); best for full MGTTA, while no-MGG baseline peaks at 70.0 with 5e-2.
  • lambda in Eqn (7) = 0.4
    Inherited from FOA; balances entropy and feature discrepancy terms.
  • Pretraining steps T and batch size = T=2000, batch size 2
    Set in Appendix B.2; not ablated.
  • phi_lr learnable vector in Eqn (5) = learned during meta-training
    Adaptive learning rate for GML updates; a learned free parameter in the memory layer.
  • Number of in-distribution samples for source statistics = 64
    From FOA; used to precompute source feature means and variances in Eqn (7).
  • MGG pretraining sample count = 128
    Chosen to show data efficiency; results are stable from N=64 to N=1000 (Table 6).
assumptions (5)
  • domain assumption The unsupervised loss in Eqn (7), combining prediction entropy and feature-statistic discrepancy, is a reliable proxy for accuracy under distribution shift.
    The entire gradient signal for both MGG training and TTA comes from this loss; if the loss is misleading, the learned gradients are misleading.
  • domain assumption An MGG pretrained on 128 unlabeled ImageNet-C validation images generalizes to unseen test streams from ImageNet-C/R/Sketch/A.
    This meta-generalization is the core transfer claim; it is tested only on ImageNet variants, not on truly disjoint domains.
  • domain assumption Source feature statistics from 64 in-distribution images remain valid for OOD test batches.
    Eqn (7) aligns test feature statistics to these precomputed source statistics; if the statistics are not representative, the adaptation objective is biased.
  • ad hoc to paper The reconstruction loss in Eqn (3) compresses historical gradient information into phi_m in a form useful for future gradient generation.
    This is the proposed mechanism; supported only by ablations, not by a formal argument.
  • ad hoc to paper Modeling each model parameter's gradient sequence independently with its own GML is sufficient (parameter-wise memorization).
    Section 3.5 assumes independence across the 38,400 updated norm-layer parameters, ignoring cross-parameter structure.
invented entities (2)
  • Meta Gradient Generator (MGG) independent evidence
    purpose: A neural-network optimizer that takes unreliable TTA gradients as input and outputs refined gradients for model updates (Eqns 2 and 6).
    The method's main contribution; its effect is evaluated via ablations (Tables 4 and 11) and it is implemented in the released code.
  • Gradient Memory Layer (GML) independent evidence
    purpose: A parameter-as-memory sequence layer that memorizes historical gradient information via reconstruction loss (Eqn 3).
    Ablated against LSTM and against no-MGG; code is available. It is a new architectural component rather than a hypothesized physical entity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning to Generate Gradients for Test-Time Adaptation via Test-Time Training Layers." pith.science (2026). https://pith.science/paper/FPDBPLV2

@misc{pith2026241216901,
  author       = {Pith},
  title        = {Pith review of: Learning to Generate Gradients for Test-Time Adaptation via Test-Time Training Layers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FPDBPLV2}},
  note         = {Machine review of arXiv:2412.16901}
}
read the original abstract

Test-time adaptation (TTA) aims to fine-tune a trained model online using unlabeled testing data to adapt to new environments or out-of-distribution data, demonstrating broad application potential in real-world scenarios. However, in this optimization process, unsupervised learning objectives like entropy minimization frequently encounter noisy learning signals. These signals produce unreliable gradients, which hinder the model ability to converge to an optimal solution quickly and introduce significant instability into the optimization process. In this paper, we seek to resolve these issues from the perspective of optimizer design. Unlike prior TTA using manually designed optimizers like SGD, we employ a learning-to-optimize approach to automatically learn an optimizer, called Meta Gradient Generator (MGG). Specifically, we aim for MGG to effectively utilize historical gradient information during the online optimization process to optimize the current model. To this end, in MGG, we design a lightweight and efficient sequence modeling layer -- gradient memory layer. It exploits a self-supervised reconstruction loss to compress historical gradient information into network parameters, thereby enabling better memorization ability over a long-term adaptation process. We only need a small number of unlabeled samples to pre-train MGG, and then the trained MGG can be deployed to process unseen samples. Promising results on ImageNet-C, R, Sketch, and A indicate that our method surpasses current state-of-the-art methods with fewer updates, less data, and significantly shorter adaptation iterations. Compared with a previous SOTA method SAR, we achieve 7.4% accuracy improvement and 4.2 times faster adaptation speed on ImageNet-C.

Figures

Figures reproduced from arXiv: 2412.16901 by the authors.

Figure 1
Figure 1. Method Differences. We devise an automatically [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. An overall illustration of MGTTA, in which we design a two-step meta gradient generator (MGG) to generate opti [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Convergence speed comparisons on ImageNet-R. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 35 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    W.; Pfau, D.; Schaul, T.; Shillingford, B.; and De Freitas, N

    Andrychowicz, M.; Denil, M.; Gomez, S.; Hoffman, M. W.; Pfau, D.; Schaul, T.; Shillingford, B.; and De Freitas, N. 2016. Learning to learn by gradient descent by gradient descent. In Advances in Neural Information Processing Systems, 3981--3989

    work page 2016

  4. [4]

    Boudiaf, M.; Mueller, R.; Ben Ayed, I.; and Bertinetto, L. 2022 a . Parameter-free online test-time adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 8344--8353

    work page 2022

  5. [5]

    Boudiaf, M.; Mueller, R.; Ben Ayed, I.; and Bertinetto, L. 2022 b . Parameter-free online test-time adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 8344--8353

    work page 2022

  6. [6]

    Chen, G.; Niu, S.; Chen, D.; Zhang, S.; Li, C.; Li, Y.; and Tan, M. 2024 a . Cross-device collaborative test-time adaptation. In Advances in Neural Information Processing Systems

    work page 2024

  7. [7]

    Chen, T.; Zhang, W.; Jingyang, Z.; Chang, S.; Liu, S.; Amini, L.; and Wang, Z. 2020. Training stronger baselines for learning to optimize. In Advances in Neural Information Processing Systems

    work page 2020

  8. [8]

    Chen, X.; Chen, T.; Cheng, Y.; Chen, W.; Awadallah, A.; and Wang, Z. 2022. Scalable learning to optimize: A learned optimizer can train big models. In Proceedings of the European Conference on Computer Vision, 389--405. Springer

    work page 2022

Show all 43 references
  1. [9]

    Chen, Y.; Niu, S.; Wang, Y.; Xu, S.; Song, H.; and Tan, M. 2024 b . Towards robust and efficient cloud-edge elastic model adaptation via selective entropy distillation. In Proceedings of the International Conference on Learning Representations

    2024

  2. [10]

    Dosovitskiy, A.; Beyer, L.; Kolesnikov, A.; Weissenborn, D.; Zhai, X.; Unterthiner, T.; Dehghani, M.; Minderer, M.; Heigold, G.; Gelly, S.; et al. 2021. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the International Conference o...

    2021

  3. [11]

    Gandelsman, Y.; Sun, Y.; Chen, X.; and Efros, A. 2022. Test-time training with masked autoencoders. In Advances in Neural Information Processing Systems, 29374--29385

    2022

  4. [12]

    Goyal, S.; Sun, M.; Raghunathan, A.; and Kolter, J. Z. 2022. Test time adaptation via conjugate pseudo-labels. In Advances in Neural Information Processing Systems, volume 35, 6204--6218

    2022

  5. [13]

    Hendrycks, D.; Basart, S.; Mu, N.; Kadavath, S.; Wang, F.; Dorundo, E.; Desai, R.; Zhu, T.; Parajuli, S.; Guo, M.; et al. 2021 a . The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference o...

    2021

  6. [14]

    Hendrycks, D.; and Dietterich, T. 2019. Benchmarking neural network robustness to common corruptions and perturbations. In Proceedings of the International Conference on Learning Representations

    2019

  7. [15]

    Hendrycks, D.; Zhao, K.; Basart, S.; Steinhardt, J.; and Song, D. 2021 b . Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 15262--15271

    2021

  8. [16]

    Hochreiter, S.; and Schmidhuber, J. 1997. Long short-term memory. Neural Computation, 9(8): 1735--1780

    1997

  9. [17]

    Iwasawa, Y.; and Matsuo, Y. 2021. Test-time classifier adjustment module for model-agnostic domain generalization. In Advances in Neural Information Processing Systems, volume 34, 2427--2440

    2021

  10. [18]

    Khurana, A.; Paul, S.; Rai, P.; Biswas, S.; and Aggarwal, G. 2021. Sita: Single image test-time adaptation. arXiv preprint arXiv:2112.02355

    2021 arXiv

  11. [19]

    P.; and Ba, J

    Kingma, D. P.; and Ba, J. 2015. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Representations

    2015

  12. [20]

    Lee, J.; Jung, D.; Lee, S.; Park, J.; Shin, J.; Hwang, U.; and Yoon, S. 2024. Entropy is not enough for test-time adaptation: From the perspective of disentangled factors. In Proceedings of the International Conference on Learning Representations

    2024

  13. [21]

    Li, C.; Chen, T.; You, H.; Wang, Z.; and Lin, Y. 2020. Halo: Hardware-aware learning to optimize. In Proceedings of the European Conference on Computer Vision, 500--518

    2020

  14. [22]

    Li, K.; and Malik, J. 2016. Learning to optimize. arXiv preprint arXiv:1606.01885

    2016 arXiv

  15. [23]

    Li, K.; and Malik, J. 2017. Learning to optimize neural nets. arXiv preprint arXiv:1703.00441

    2017 arXiv

  16. [24]

    Liu, Y.; Kothari, P.; Van Delft, B.; Bellot-Gurlet, B.; Mordan, T.; and Alahi, A. 2021. TTT++: When does self-supervised test-time training fail or thrive? In Advances in Neural Information Processing Systems, 21808--21820

    2021

  17. [25]

    Lv, K.; Jiang, S.; and Li, J. 2017. Learning gradient descent: Better generalization and longer horizons. In Proceedings of the International Conference on Machine Learning, 2247--2255

    2017

  18. [26]

    Metz, L.; Maheswaranathan, N.; Nixon, J.; Freeman, D.; and Sohl-Dickstein, J. 2019. Understanding and correcting pathologies in the training of learned optimizers. In Proceedings of the International Conference on Machine Learning, 4556--4565

    2019

  19. [27]

    J.; Micorek, J.; Possegger, H.; and Bischof, H

    Mirza, M. J.; Micorek, J.; Possegger, H.; and Bischof, H. 2022. The norm must go on: dynamic unsupervised domain adaptation by normalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 14765--14775

    2022

  20. [28]

    Nado, Z.; Padhy, S.; Sculley, D.; D'Amour, A.; Lakshminarayanan, B.; and Snoek, J. 2020. Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963

    2020 arXiv

  21. [29]

    Niu, S.; Miao, C.; Chen, G.; Wu, P.; and Zhao, P. 2024. Test-time model adaptation with only forward passes. In Proceedings of the International Conference on Machine Learning

    2024

  22. [30]

    Niu, S.; Wu, J.; Zhang, Y.; Chen, Y.; Zheng, S.; Zhao, P.; and Tan, M. 2022. Efficient test-time model adaptation without forgetting. In Proceedings of the International Conference on Machine Learning, 16888--16905

    2022

  23. [31]

    Niu, S.; Wu, J.; Zhang, Y.; Wen, Z.; Chen, Y.; Zhao, P.; and Tan, M. 2023. Towards stable test-time adaptation in dynamic wild world. In Proceedings of the International Conference on Learning Representations

    2023

  24. [32]

    Sun, Y.; Li, X.; Dalal, K.; Xu, J.; Vikram, A.; Zhang, G.; Dubois, Y.; Chen, X.; Wang, X.; Koyejo, S.; et al. 2024. Learning to (learn at test time): Rnns with expressive hidden states. arXiv preprint arXiv:2407.04620

    2024 arXiv

  25. [33]

    Sun, Y.; Wang, X.; Liu, Z.; Miller, J.; Efros, A.; and Hardt, M. 2020. Test-time training with self-supervision for generalization under distribution shifts. In Proceedings of the International Conference on Machine Learning, 9229--9248

    2020

  26. [34]

    Tan, M.; Chen, G.; Wu, J.; Zhang, Y.; Chen, Y.; Zhao, P.; and Niu, S. 2024. Uncertainty-calibrated test-time model adaptation without forgetting. arXiv preprint arXiv:2403.11491

    2024 arXiv

  27. [35]

    Vicol, P.; Metz, L.; and Sohl-Dickstein, J. 2021. Unbiased gradient estimation in unrolled computation graphs with persistent evolution strategies. In Proceedings of the International Conference on Machine Learning, 10553--10563. PMLR

    2021

  28. [36]

    Wang, D.; Shelhamer, E.; Liu, S.; Olshausen, B.; and Darrell, T. 2021. Tent: Fully test-time adaptation by entropy minimization. In Proceedings of the International Conference on Learning Representations

    2021

  29. [37]

    Wang, H.; Ge, S.; Lipton, Z.; and Xing, E. P. 2019. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, 10506--10518

    2019

  30. [38]

    Wang, Q.; Fink, O.; Van Gool, L.; and Dai, D. 2022. Continual test-time domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 7201--7211

    2022

  31. [39]

    Wen, Z.; Niu, S.; Li, G.; Wu, Q.; Tan, M.; and Wu, Q. 2023. Test-time model adaptation for visual question answering with debiased self-supervisions. IEEE Transactions on Multimedia

    2023

  32. [40]

    W.; Colmenarejo, S

    Wichrowska, O.; Maheswaranathan, N.; Hoffman, M. W.; Colmenarejo, S. G.; Denil, M.; Freitas, N.; and Sohl-Dickstein, J. 2017. Learned optimizers that scale and generalize. In Proceedings of the International Conference on Machine Learning, 3751--3760

    2017

  33. [41]

    Yuan, L.; Xie, B.; and Li, S. 2023. Robust test-time adaptation in dynamic scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 15922--15932

    2023

  34. [42]

    Zeng, R.; Deng, Q.; Xu, H.; Niu, S.; and Chen, J. 2023. Exploring motion cues for video test-time adaptation. In Proceedings of the International Conference on Multimedia, 1840--1850

    2023

  35. [43]

    Zhang, M.; Levine, S.; and Finn, C. 2022. Memo: Test time robustness via adaptation and augmentation. In Advances in Neural Information Processing Systems, 38629--38642

    2022

Pith tools

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