Pith. sign in

REVIEW 4 major objections 5 minor 35 references

Feature-Based Instance Neighbor Discovery: Advanced Stable Test-Time Adaptation in Dynamic World

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read FIND keeps test-time adaptation working when a batch mixes several distributions: it partitions each layer's features into similar groups, normalizes each with blended statistics, and reports 30% accuracy gains over prior methods.

desk verdict Plausible normalization method and useful dynamic scenarios; the 30% claim overreaches and cluster purity is untested. read the letter →

arxiv 2506.06782 v1 pith:2UWUSCH2 submitted 2025-06-07 cs.LG cs.AIcs.CV

Qinting Jiang , Chuyang Ye , Dongyan Wei , Bingli Wang , Yuan Xue , Jingyan Jiang , Zhi Wang This is my paper
classification cs.LGcs.AIcs.CV
keywords test-timeadaptationbatchnormalizationdistributionshiftdynamicscenariosmulti-distributionbatchesfirst-neighborclusteringlayer-wisecorruptionrobustness
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Test-time adaptation lets a deployed model fix itself as it sees new data, but existing methods assume the samples in one batch come from a single distribution. Real streams can mix several at once, and then computing one global mean and variance for the whole batch conflates the distributions and costs accuracy. This paper claims the fix is a divide-and-conquer normalization: at each batch-normalization layer, group the batch's feature maps by similarity using a parameter-free first-neighbor clustering of per-sample channel means, normalize each group with its own statistics, and blend those with the source model's statistics so small groups keep generic class knowledge. In dynamic test scenarios on corrupted CIFAR and ImageNet benchmarks, the method reports roughly 30% higher accuracy than prior test-time adaptation methods, with no gradient updates, low memory and latency, and stable behavior across batch sizes and architectures including a transformer.

What carries the argument

Three mechanisms carry the argument. LFD (Layer-wise Feature Disentanglement): at a BN layer each sample is represented by its instance-level channel-mean vector mu_I in R^C (Eq. 1); cosine similarity between these vectors (Eq. 2) picks each sample's first neighbor, and the adjacency rule that connects samples sharing a first neighbor (Eq. 3) groups the batch into connected components, a parameter-free clustering taken from FINCH. FABN (Feature Aware Batch Normalization): per-group statistics mu_F, sigma_F (Eq. 4) are blended with source statistics through $\alpha$-weighting (Eqs. 5-6) before group-wise normalization (Eq. 7). S-FABN (Selective FABN): a per-layer sensitivity score combines the mean and standard deviation of channel-wise KL divergence between target and source (Eqs. 8-9); layers below a threshold gamma skip partitioning and fall back to whole-batch test statistics.

What would settle it

Run FIND in the CrossMix scenario with ground-truth corruption labels available and measure, layer by layer, how often an LFD cluster contains samples from only one corruption type (purity or adjusted Rand index against the domain labels). If the layers that carry most of the reported accuracy gain produce clusters no purer than random partitions of the same size while accuracy is maintained, the claimed mechanism of distribution-coherent grouping is not what causes the gain; a control experiment that replaces cosine-similarity grouping with random splits of identical cluster sizes would settle this.

Watch

Extended reading notes

Core claim

FIND's central claim is that the failure of test-time batch normalization in dynamic scenarios comes from computing statistics globally over a mixture of distributions, and that the mixture structure can be recovered, label-free, from the features themselves. At each BN layer, per-sample channel-mean vectors form distinct clusters with different means and variances, so LFD groups samples through first-neighbor relations in cosine-similarity space, producing subsets whose test-feature-specific statistics describe one distribution rather than a blend. FABN then mixes each subset's statistics with the source batch-normalization statistics at a fixed weight (alpha = 0.8), which the paper argues compensates for the generic information that small subsets lose. S-FABN derives a per-layer sensitivity score from the mean and spread of the KL divergence between target and source statistics, and switches partitioning off for insensitive layers, so the selective variant FIND* runs faster at nearly equal accuracy. Across CIFAR10-C, CIFAR100-C and ImageNet-C under three dynamic protocols, the paper reports accuracy gains of roughly 30% over prior state-of-the-art, plus the first strong result for a BN-based test-time normalization method on a transformer backbone.

Load-bearing premise

The method stands on the assumption that grouping a batch's samples by how similar their per-channel average features are at each layer yields groups that each come from a single distribution; if a group mixes several distributions, its normalization statistics are biased and the source-statistics blend is only a rough correction.

Editorial extensions

If this is right

  • Test-time normalization can be made robust to batches that mix many distributions at once: the paper reports accuracy gains of about 30% over the best prior methods across the CrossMix, Random, and Shuffle scenarios.
  • Because FIND only adjusts normalization statistics and never backpropagates, it stays cheap, with around 0.21 s latency and 1.44 GB memory on ImageNet-C with ResNet-50, tens of times faster than some fine-tuning baselines.
  • The selective variant FIND* shows that many layers do not need partitioning at all; turning it off cuts runtime further (0.15 s) with negligible accuracy loss, which helps when streams are long.
  • Accuracy stays roughly constant as batch size shrinks from 64 down to 1, a regime where earlier normalization baselines degrade, which matters for small-batch and on-device deployment.
  • The same normalization scheme transfers to a transformer-style backbone with BN layers (EfficientViT), supporting the paper's claim of being the first BN-based test-time adaptation method that works well on such models.

Reading between the lines

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

  • If LFD's grouping is really the source of the gain, the same instance-level channel-mean descriptor could be used to harden fine-tuning-based test-time adaptation against mixed batches, for example by restricting gradient updates to samples from one detected group at a time.
  • The fixed blend weight alpha = 0.8 and the threshold gamma are the method's free parameters; the paper's own per-layer sensitivity score suggests a natural extension in which alpha is set per layer, which the authors themselves list as future work.
  • The three dynamic protocols (CrossMix, Random, Shuffle) form a reusable stress test: any test-time adaptation method could be evaluated on mixed-distribution batches with abrupt shifts before deployment, since the paper shows prior methods lose over 15% accuracy on average in this setting.
  • A direct check of the mechanism would be measuring the domain purity of LFD's clusters against ground-truth corruption labels; the paper's own Figure 12 indicates clusters track domains in middle layers and classes in deep layers, implying the benefit of partitioning is concentrated in the middle layers.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The manuscript proposes FIND, a test-time normalization method for dynamic scenarios in which test batches contain samples from multiple distributions. The method has three components: Layer-wise Feature Disentanglement (LFD), which partitions per-layer feature maps by first-neighbor relationships among instance-level channel means; Feature Aware Batch Normalization (FABN), which blends per-cluster test statistics with source batch-normalization statistics; and Selective FABN (S-FABN), which disables partitioning on layers judged insensitive by a KL-divergence-based score. The authors evaluate FIND on CIFAR10-C, CIFAR100-C, and ImageNet-C under CrossMix, Random, and Shuffle scenarios with ResNet-50 and EfficientViT backbones, and report accuracy and efficiency comparisons against TENT, EATA, NOTE, SAR, RoTTA, ViDA, DeYO, TBN, alpha-BN, and IABN. The central claim is that FIND achieves a 30% accuracy improvement in dynamic scenarios, attributed to the divide-and-conquer normalization strategy.

Significance. If the reported gains are robust, the paper addresses a real and under-explored problem: test-time normalization when batches contain multiple simultaneous distributions. The experimental breadth is a genuine strength: three corrupted benchmark datasets, three dynamic scenario definitions, two backbones, a long-horizon replication study, and an efficiency analysis. The method is also backward-free, which is attractive for deployment. However, the paper currently does not provide the evidence needed to support its central mechanism: LFD's cluster purity is never measured, the headline 30% figure is not supported by the aggregate tables, and key hyperparameters are tuned on the evaluation datasets. These issues are fixable in principle, but they are load-bearing for the paper's claims.

major comments (4)
  1. [Abstract and Section 3.2, Table 1] The abstract's claim of a 30% accuracy improvement in dynamic scenarios is not supported by Table 1. In the Avg-All column, FIND reaches 47.87% and FIND* 48.50%, while the strongest baseline alpha-BN reaches 43.98% and DeYO reaches 43.48%. The gain over these baselines is about 4-5 percentage points, i.e., roughly 9-11% relative, not 30%. The largest single-cell gain in Table 1 is on ImageNet-C under CrossMix (FIND 29.21% vs DeYO 19.13%), but this is one favorable cell, not an aggregate result. The authors should either correct the claim or specify precisely which comparison yields 30%.
  2. [Sections 2.3-2.4 and Appendix H] The load-bearing premise of the method is that LFD partitions each batch into distributionally coherent clusters, so that the per-cluster TFN statistics in Eq. (4) are single-distribution estimates. This premise is not demonstrated. Appendix H/Fig. 12 only reports the number of clusters per layer and observes that the count matches the number of domains in middle layers; a match of counts does not establish purity, since heavily mixed clusters can produce the same count. No ARI, NMI, or per-cluster domain-composition statistics are reported. The scale issue compounds this: in CrossMix with batch size 64 and 15 domains, an average cluster contains about 4 samples, so even a pure cluster yields noisy mean and variance estimates in Eq. (4). The authors should report direct cluster-purity metrics and, ideally, compare FABN with an oracle domain-partition baseline.
  3. [Section 3.1 and Appendices E-G] Key hyperparameters appear to be selected using the evaluation datasets. The aggregation weight alpha is set to 0.8, the S-FABN threshold gamma is set to 0.1, and the cold-start duration is set to 10 batches based on sensitivity analyses on CIFAR10-C, CIFAR100-C, and ImageNet-C (Section 3.1, Figures 6 and 11, Appendix F). No held-out validation set or cross-validation procedure is described. Because these choices are tuned on the test domains, the reported numbers include fitted hyperparameters. The authors should either use fixed values justified by a separate validation split or report results over a range of hyperparameter choices in a way that does not require access to the test distribution.
  4. [Section 3.5, Table 3] The clustering ablation in Table 3 does not isolate the contribution of LFD's specific grouping mechanism. It compares LFD against HDBSCAN, Agglomerative, Birch, RccCluster, DSets-DBSCAN, DBSCAN, and K-means, but it does not include a size-matched random partition or a global (non-partitioned) baseline using the same number of clusters. Without such a control, the improvement could come from the mere act of splitting the batch into small groups rather than from the distributional coherence of LFD's first-neighbor groups. A random-partition or global-partition ablation is needed to support the divide-and-conquer attribution.
minor comments (5)
  1. [Section 2.4] In the text after Eq. (3), the set F is written as {F^1, F^2, ..., F^2, ..., F^r}, which contains a duplicated F^2; this should be {F^1, F^2, ..., F^r}.
  2. [Section 3.1] EfficientViT is described as a vision transformer, but it is a hybrid architecture with both convolution and attention components; the claim of being 'the first BN-based method tested on ViT' should be stated more precisely or softened.
  3. [Appendix C] The baseline descriptions contain formatting artifacts such as 'TEST-T IME FINE -T UNE' and inconsistent bullet indentation; these should be cleaned up.
  4. [Appendix K] The description of Figure 13 says the analysis computes distance between 'instance-normalized and TBN statistics,' but the text is unclear about whether the comparison is between per-instance statistics and batch statistics or between two normalization outputs; please clarify the exact quantity plotted.
  5. [Section 5 and Appendix P] The conclusion and Appendix P contain duplicated sentences ('Our approach demonstrates robust performance across diverse distribution shifts...' appears twice in Section 5), and the text should be edited for conciseness.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity found: the method is an empirical proposal validated on external benchmarks, not a derivation that reduces to its inputs.

full rationale

The paper proposes an algorithmic method (LFD clustering plus FABN statistics blending) and evaluates it on external benchmarks (CIFAR10-C, CIFAR100-C, ImageNet-C under the TTAB protocol). I looked for self-definitional reductions: Eq. (4)-(7) compute per-subset normalization statistics and blend them with source statistics via alpha; although the subsets are formed from per-instance channel means in Eqs. (1)-(3), the FABN output is not equal to those inputs by construction, because it depends on the full spatial feature maps, subset statistics, and affine parameters. There is no fitted parameter that is renamed as a prediction: alpha=0.8 and gamma=0.1 are hyperparameters chosen through sensitivity experiments, and the reported accuracies are measured results under those settings rather than predictions forced by the equations. The paper cites external building blocks (FINCH, TTAB, Neyshabur et al., Lee et al.) rather than invoking the authors' own prior results as a uniqueness theorem. The empirical weakness that cluster purity is not directly measured in Appendix H is a validity concern, not a circularity: a cluster count matching the domain count does not prove domain-coherent subsets, but this is a missing measurement, not a demonstration that the derivation reduces to its inputs. No load-bearing self-citation or definitional equivalence was found.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on several domain assumptions about cluster separability and the validity of source statistics under shift, plus three fitted hyperparameters. No new physical or conceptual entities are introduced; the FIND modules are compositions of known operations.

free parameters (3)
  • alpha (FABN blending weight) = 0.8
    Weight in Eq. 5-6 controlling SBN versus TFN contributions; set to 0.8 in Section 3.1 and tuned on CIFAR/ImageNet-C. Sensitivity analysis in Appendix G shows accuracy is not very sensitive in a narrow range, but the selected value is an empirical choice on the same benchmarks.
  • gamma (S-FABN threshold) = 0.1
    Layer-selection threshold in Section 2.5; accuracy drops for gamma above 0.1 (Figure 6 and Figure 11). The selected value is fitted to the evaluation data, with a sensitivity analysis on the same datasets.
  • cold-start duration = 10 batches
    Number of initial batches used to compute per-layer sensitivity scores in S-FABN (Appendix F). The paper reports negligible impact across 3 to 100 batches and chooses 10.
assumptions (4)
  • domain assumption Feature distributions of different domains in a test batch form distinct clusters at each BN layer, and per-instance channel means are sufficient statistics to separate them.
    Motivation in Section 2.2 and Figure 3a; LFD (Section 2.3) partitions based on instance-wise means mu_I using cosine similarity (Eq. 1-3). If cluster boundaries do not match domain or class groups, FABN computes statistics on mixed distributions.
  • domain assumption The first-neighbor graph partitioning (FINCH-style, from [23]) yields useful clusters without a predefined cluster count.
    Section 2.3 uses connected components of the first-neighbor adjacency matrix (Eq. 3). There is no theoretical guarantee that this partitions feature maps into distributionally homogeneous subsets.
  • domain assumption Source BN statistics (SBN) remain valid under distribution shift and improve stability of cluster-level normalization (FABN, Eq. 5-6).
    Section 2.4 assumes that combining SBN with test cluster statistics reduces bias from small clusters; this is plausible but not proven.
  • domain assumption Layer-wise KL scores (Eq. 8-9) correctly rank layer sensitivity to distribution shift, so disabling partitioning for low-scoring layers does not hurt accuracy.
    S-FABN's selection rule (Section 2.5) relies on this; the empirical support is the gamma sensitivity analysis in Section 3.4 and Appendix G.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Feature-Based Instance Neighbor Discovery: Advanced Stable Test-Time Adaptation in Dynamic World." pith.science (2026). https://pith.science/paper/2UWUSCH2

@misc{pith2026250606782,
  author       = {Pith},
  title        = {Pith review of: Feature-Based Instance Neighbor Discovery: Advanced Stable Test-Time Adaptation in Dynamic World},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2UWUSCH2}},
  note         = {Machine review of arXiv:2506.06782}
}
read the original abstract

Despite progress, deep neural networks still suffer performance declines under distribution shifts between training and test domains, leading to a substantial decrease in Quality of Experience (QoE) for applications. Existing test-time adaptation (TTA) methods are challenged by dynamic, multiple test distributions within batches. We observe that feature distributions across different domains inherently cluster into distinct groups with varying means and variances. This divergence reveals a critical limitation of previous global normalization strategies in TTA, which inevitably distort the original data characteristics. Based on this insight, we propose Feature-based Instance Neighbor Discovery (FIND), which comprises three key components: Layer-wise Feature Disentanglement (LFD), Feature Aware Batch Normalization (FABN) and Selective FABN (S-FABN). LFD stably captures features with similar distributions at each layer by constructing graph structures. While FABN optimally combines source statistics with test-time distribution specific statistics for robust feature representation. Finally, S-FABN determines which layers require feature partitioning and which can remain unified, thereby enhancing inference efficiency. Extensive experiments demonstrate that FIND significantly outperforms existing methods, achieving a 30\% accuracy improvement in dynamic scenarios while maintaining computational efficiency.

Figures

Figures reproduced from arXiv: 2506.06782 by the authors.

Figure 1
Figure 1. (a) In previous TBN-based methods, batch-level statistical values are computed holisti￾cally across all features, leading to erroneous nor￾malization of cross-distribution features. (b) Our FIND framework introduces a layer-wise divide￾and-conquer strategy for intra-batch feature pro￾cessing, synergistically integrating source domain knowledge to achieve reliable normalization. This brings us to a crucial insight: i… view at source ↗
Figure 2
Figure 2. (a) demonstrates the performance gains of test-time normalization versus test-time fine￾tuning (FT). (b) shows how gradient changes across fine-tuning steps under different scenarios. A divide-and-conquer strategy is essential in dynamic environments. Figure 3a re￾veals intrinsic cluster patterns in BN lay￾ers when using domain as labels, where each cluster exhibits unique distribution char￾acteristics (varying cent… view at source ↗
Figure 3
Figure 3. Feature distributions across domains (a) and classes (b and c). Replace processing the input samples directly by layer￾wise feature partitioning. Direct partitioning input sam￾ples is infeasible due to missing domain labels and domain￾class feature couplings. Inspired by Neyshabur et al. [18] and Lee et al. [11], who demonstrated that visual features inherently decompose into domain-relevant features (DRF, e.g., bac… view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: (a) and (b) display the inference accuracy [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Overview of the proposed FIND (FIND*). We design layer-wise feature disentanglement (LFD) to divide multi-distribution feature maps within each layer into different subsets. Features in each subset have similar cosine distances, ensuring distribution consistency. Next,…
Figure 6
Figure 6. Figure 6: (a) and (b) demonstrate the sensitivity of model inference performance (Accuracy (%)) to γ variations, along with the associated changes in inference efficiency (Time (s)). 8 [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Batch size sensitivity analysis. Our approach [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Visualization of the 15 types of corruptions in CIFAR10-C, CIFAR100-C and ImageNet-C. [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: The visualization of scenarios mentioned in this paper, including static and dynamic [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]
Figure 10
Figure 10. Figure 10: (a)-(c) demonstrate the variation trends of domain shift sensitivity across BN layers under different backbones and datasets. (d) shows the sensitivity variations across BN layers in a single model under various cold-start durations. compromise the final layer selecti…
Figure 11
Figure 11. Figure 11: (a) and (b) demonstrate the sensitivity of model inference performance (Accuracy (%)) to [PITH_FULL_IMAGE:figures/full_fig_p018_11.png]
Figure 12
Figure 12. Figure 12: Cluster numbers of each FABN layer under dynamic (Random, CrossMix, and Shuffle) scenarios. Each result is derived from the mean of 1,000 batches, with the col￾ored bands around the lines representing the standard deviation. Our analysis reveals the fundamental mechan…
Figure 13
Figure 13. Figure 13: TBN-IN distance vs. accuracy under different corruptions. The 5 data points in the [PITH_FULL_IMAGE:figures/full_fig_p020_13.png]
Figure 14
Figure 14. Figure 14: Airplane wings in different corruption. More severe corruption perturbations of the image [PITH_FULL_IMAGE:figures/full_fig_p020_14.png]
Figure 15
Figure 15. Figure 15: Batch size vs. accuracy. Other methods exhibit poorer performance at low batch sizes, [PITH_FULL_IMAGE:figures/full_fig_p021_15.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 28 canonical work pages

  1. [1]

    The k-means algorithm: A comprehensive survey and performance evaluation

    Mohiuddin Ahmed, Raihan Seraj, and Syed Mohammed Shamsul Islam. The k-means algorithm: A comprehensive survey and performance evaluation. Electronics, 9(8):1295, 2020

    work page 2020

  2. [2]

    Tasknorm: Rethinking batch normalization for meta-learning

    John Bronskill, Jonathan Gordon, James Requeima, Sebastian Nowozin, and Richard Turner. Tasknorm: Rethinking batch normalization for meta-learning. In International Conference on Machine Learning, pages 1153–1164. PMLR, 2020

    work page 2020

  3. [3]

    Contrastive Test-Time Adapta- tion, April 2022

    Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive Test-Time Adapta- tion, April 2022. arXiv:2204.10377 [cs]

    arXiv 2022

  4. [4]

    Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening

    Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11580–11590, 2021

    work page 2021

  5. [5]

    Back to the Source: Diffusion-Driven Test-Time Adaptation, June 2023

    Jin Gao, Jialing Zhang, Xihui Liu, Trevor Darrell, Evan Shelhamer, and Dequan Wang. Back to the Source: Diffusion-Driven Test-Time Adaptation, June 2023. arXiv:2207.03442 [cs]

    arXiv 2023

  6. [6]

    NOTE: Robust Continual Test-time Adaptation Against Temporal Correlation, 2023

    Taesik Gong, Jongheon Jeong, Taewon Kim, Yewon Kim, Jinwoo Shin, and Sung-Ju Lee. NOTE: Robust Continual Test-time Adaptation Against Temporal Correlation, 2023

    work page 2023

  7. [7]

    Deep Residual Learning for Image Recognition, 2015

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition, 2015

    2015

  8. [8]

    Benchmarking neural network robustness to common corruptions and perturbations, 2019

    Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations, 2019

    2019

Show all 35 references
  1. [9]

    Dsets-dbscan: A parameter-free clustering algorithm

    Jian Hou, Huijun Gao, and Xuelong Li. Dsets-dbscan: A parameter-free clustering algorithm. IEEE Transactions on Image Processing, 25(7):3182–3193, 2016

    2016

  2. [10]

    Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift, 2015

    Sergey Ioffe and Christian Szegedy. Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift, 2015

    2015

  3. [11]

    Entropy is not enough for test-time adaptation: From the perspective of disentangled factors, 2024

    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 perspective of disentangled factors, 2024

    2024

  4. [12]

    TTN: A Domain-Shift Aware Batch Normalization in Test-Time Adaptation, 2023

    Hyesu Lim, Byeonggeun Kim, Jaegul Choo, and Sungha Choi. TTN: A Domain-Shift Aware Batch Normalization in Test-Time Adaptation, 2023

    2023

  5. [13]

    Vida: Homeostatic visual domain adapter for continual test time adaptation, 2024

    Jiaming Liu, Senqiao Yang, Peidong Jia, Renrui Zhang, Ming Lu, Yandong Guo, Wei Xue, and Shanghang Zhang. Vida: Homeostatic visual domain adapter for continual test time adaptation, 2024

    2024

  6. [14]

    Efficientvit: Memory efficient vision transformer with cascaded group attention, 2023

    Xinyu Liu, Houwen Peng, Ningxin Zheng, Yuqing Yang, Han Hu, and Yixuan Yuan. Efficientvit: Memory efficient vision transformer with cascaded group attention, 2023

    2023

  7. [15]

    hdbscan: Hierarchical density based clustering

    Leland McInnes, John Healy, Steve Astels, et al. hdbscan: Hierarchical density based clustering. J. Open Source Softw., 2(11):205, 2017

    2017

  8. [16]

    The norm must go on: Dynamic unsupervised domain adaptation by normalization

    M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsupervised domain adaptation by normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14765–14775, 2022

    2022

  9. [17]

    Ward’s hierarchical agglomerative clustering method: which algorithms implement ward’s criterion? Journal of classification, 31:274–295, 2014

    Fionn Murtagh and Pierre Legendre. Ward’s hierarchical agglomerative clustering method: which algorithms implement ward’s criterion? Journal of classification, 31:274–295, 2014

    2014

  10. [18]

    What is being transferred in transfer learning? Advances in neural information processing systems, 33:512–523, 2020

    Behnam Neyshabur, Hanie Sedghi, and Chiyuan Zhang. What is being transferred in transfer learning? Advances in neural information processing systems, 33:512–523, 2020

    2020

  11. [19]

    Efficient test-time model adaptation without forgetting, 2022

    Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting, 2022

    2022

  12. [20]

    Towards stable test-time adaptation in dynamic wild world, 2023

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

    2023

  13. [21]

    Dataset shift in machine learning

    Joaquin Quiñonero-Candela, Masashi Sugiyama, Anton Schwaighofer, and Neil D Lawrence. Dataset shift in machine learning. Mit Press, 2022

    2022

  14. [22]

    Density-based clustering in spatial databases: The algorithm gdbscan and its applications

    Jörg Sander, Martin Ester, Hans-Peter Kriegel, and Xiaowei Xu. Density-based clustering in spatial databases: The algorithm gdbscan and its applications. Data mining and knowledge discovery, 2:169–194, 1998. 11

    1998

  15. [23]

    Efficient parameter-free clustering using first neighbor relations

    Saquib Sarfraz, Vivek Sharma, and Rainer Stiefelhagen. Efficient parameter-free clustering using first neighbor relations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8934–8943, 2019

    2019

  16. [24]

    Improving robustness against common corruptions by covariate shift adaptation

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

    2020

  17. [25]

    Recursive consensus clustering for novel subtype discovery from transcriptome data

    Pranali Sonpatki and Nameeta Shah. Recursive consensus clustering for novel subtype discovery from transcriptome data. Scientific reports, 10(1):11005, 2020

    2020

  18. [26]

    Tent: Fully Test-time Adaptation by Entropy Minimization, 2021

    Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully Test-time Adaptation by Entropy Minimization, 2021

    2021

  19. [27]

    Decoupled Prototype Learning for Reliable Test-Time Adaptation, January 2024

    Guowei Wang, Changxing Ding, Wentao Tan, and Mingkui Tan. Decoupled Prototype Learning for Reliable Test-Time Adaptation, January 2024. arXiv:2401.08703 [cs]

    2024 arXiv

  20. [28]

    Continual Test-Time Domain Adapta- tion, 2022

    Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual Test-Time Domain Adapta- tion, 2022

    2022

  21. [29]

    Dynamically instance-guided adaptation: A backward-free approach for test-time domain adaptive semantic segmentation

    Wei Wang, Zhun Zhong, Weijie Wang, Xi Chen, Charles Ling, Boyu Wang, and Nicu Sebe. Dynamically instance-guided adaptation: A backward-free approach for test-time domain adaptive semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Re...

    2023

  22. [30]

    Test-time batch statistics calibration for covariate shift

    Fuming You, Jingjing Li, and Zhou Zhao. Test-time batch statistics calibration for covariate shift. arXiv preprint arXiv:2110.04065, 2021

    2021 arXiv

  23. [31]

    Robust Test-Time Adaptation in Dynamic Scenarios, 2023

    Longhui Yuan, Binhui Xie, and Shuang Li. Robust Test-Time Adaptation in Dynamic Scenarios, 2023

    2023

  24. [32]

    Memo: Test time robustness via adaptation and augmentation

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

    2022

  25. [33]

    Birch: an efficient data clustering method for very large databases

    Tian Zhang, Raghu Ramakrishnan, and Miron Livny. Birch: an efficient data clustering method for very large databases. ACM sigmod record, 25(2):103–114, 1996

    1996

  26. [34]

    On pitfalls of test-time adaptation

    Hao Zhao, Yuejiang Liu, Alexandre Alahi, and Tao Lin. On pitfalls of test-time adaptation. In International Conference on Machine Learning (ICML), 2023

    2023

  27. [35]

    Xingzhi Zhou, Zhiliang Tian, Ka Chun Cheung, Simon See, and Nevin L. Zhang. Resilient Practical Test-Time Adaptation: Soft Batch Normalization Alignment and Entropy-driven Memory Bank, January 2024. arXiv:2401.14619 [cs]. 12 Contents of Appendix A The Layer-Wise Feature Disent...

    2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.