Pith. sign in

REVIEW 3 major objections 5 minor 106 references

Reliable Few-shot Learning under Dual Noises

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A plug-in task-adaptation module filters both in-distribution background clutter and out-of-distribution samples from support and query sets, improving few-shot classification accuracy and cutting false-positive rates in few-shot OOD…

desk verdict A substantial empirical extension of DETA with real plug-in value, but the core CoRA statistic is undefined for 1-shot classes, which undermines the Table 3 state-of-the-art claims until the authors specify a fallback. read the letter →

arxiv 2506.16330 v1 pith:ELBEGQI4 submitted 2025-06-19 cs.CV cs.AI

Ji Zhang , Jingkuan Song , Lianli Gao , Nicu Sebe , Heng Tao Shen This is my paper
classification cs.CVcs.AI
keywords few-shotlearningout-of-distributiondetectiontaskadaptationdatadenoisingcontrastiverelevanceaggregationprototypelossmemorybankprompttuning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that few-shot models trained by task adaptation break down in the open world because two kinds of noise invade the support and query sets of every task: in-distribution noise, meaning cluttered backgrounds attached to valid class images, and out-of-distribution noise, meaning whole images from classes the model has never seen. With only a handful of support samples, a few noisy images can dominate adaptation. DETA++ is a proposed plugin layer that weighs each cropped image region by how strongly it coheres to same-class regions versus other-class regions, down-weights noisy regions and whole noisy images, trains a clean prototype loss and a noise entropy-maximization loss, and at inference classifies by nearest centroid over clean regions held in a memory bank. The paper reports that attaching DETA++ lifts accuracy across six task-adaptation baselines on Meta-Dataset and cuts the average FPR95 (false positive rate of out-of-distribution samples when 95 percent of in-distribution samples are accepted) of a strong prompt-tuning OOD detection baseline from 39.48 to 30.88 on ImageNet-1K.

What carries the argument

CoRA (Contrastive Relevance Aggregation): for each cropped region, the weight is $\lambda_{ij} = \tilde{\varphi}(z_{ij})/\tilde{\psi}(z_{ij})$, where $\tilde{\varphi}$ is the normalized mean cosine similarity to in-class regions (excluding the image's own other regions, to avoid domination) and $\tilde{\psi}$ is the normalized mean cosine similarity to out-of-class regions. This single ratio carries the argument: it yields region weights that filter in-distribution noise (background clutter) by low in-class relevance, image weights that filter out-of-distribution noise by momentum-averaging region weights, the split into clean set and noisy set under threshold $\varrho$, the clean prototype loss, the noise entropy-maximization loss, and the memory bank of top-$2K$ clean regions that feed IntraSwap and LocalNCC. The total adaptation objective is $\mathcal{L} = \mathcal{L}_{\mathrm{clean}} + \beta \mathcal{L}_{\mathrm{noise}}$.

What would settle it

Create or annotate a few-shot task where every region's clean/noisy status is known, for instance by pasting synthetic background patches or other-class objects into support images, then measure whether CoRA's region weights with the reported threshold value separate the two sets; if a large fraction of clean object regions fall below the threshold or noisy background regions lie above it, the clean prototype loss and LocalNCC would be training and evaluating on the wrong regions, contradicting the mechanism.

Watch

Extended reading notes

Core claim

The central claim DETA++ argues for is that the same local-region statistics can denoise both support and query samples for both noise types. CoRA computes, for every randomly cropped region, its mean cosine similarity to regions of the same class and to regions of other classes; the ratio of normalized in-class over out-of-class scores becomes the region's weight. A threshold splits regions into clean and noisy. Clean regions define weighted class prototypes and drive the clean prototype loss, pulling clean features close to their prototypes; noisy regions drive the noise entropy-maximization loss, which maximizes the entropy of predictions so noise features are pushed off the decision boundary. A memory bank keeps the highest-weight clean regions per class, IntraSwap pastes those regions back into same-class images to rectify biased prototypes, and at inference LocalNCC classifies each query by its nearest class centroid in a low-dimensional embedding. The paper's claim is that this unified pipeline improves both few-shot classification under dual noise and few-shot OOD detection, and that the improvements transfer across adapter-finetuning, full-finetuning, and prompt-tuning baselines.

Load-bearing premise

Everything rests on the assumption that one fixed threshold separates clean from noisy regions and images for every class, dataset, and noise level; a single scalar weight ratio cannot be guaranteed to cleanly separate the two populations, and if the threshold is wrong, the method either discards informative object regions or keeps the noisy regions it was meant to remove.

Editorial extensions

If this is right

  • Attaching DETA++ to a task-adaptation baseline is claimed to improve few-shot classification accuracy on vanilla Meta-Dataset by 2.1 to 5.2 percentage points and on OOD-polluted Meta-Dataset by 2.7 to 6.6 percentage points over the baselines.
  • For few-shot OOD detection with prompt tuning, DETA++ is claimed to cut the average FPR95 from 39.48 to 30.88 on ImageNet-1K across iNaturalist, SUN, Places365, Texture, and NINCO.
  • Because DETA++ operates on region weights and losses rather than backbone weights, the same plugin should transfer across adapter-based, full-finetuning, and prompt-tuning task adaptation, requiring only a projection head.
  • The two proposed losses are complementary: the clean prototype loss improves classification by denoising prototypes, while the noise entropy-maximization loss improves OOD detection by pushing detected noisy regions away from all class prototypes; ablation results attribute most of the OOD gain to the noise entropy loss.
  • LocalNCC built from clean regions in the memory bank is claimed to outperform a global nearest-centroid classifier, so denoised inference matters beyond denoised adaptation.

Reading between the lines

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

  • Because CoRA is parameter-free and uses only support-set statistics, the same region-weighting could be dropped into other low-label regimes such as semi-supervised test-time adaptation, where unlabeled batches contain corrupt or cross-domain regions.
  • The paper uses a single scalar threshold tuned per benchmark; an adaptive threshold fit per task, for instance from the weight distribution's quantiles, would be a natural extension and would test whether the clean/noisy separation is truly bimodal.
  • The entropy-maximization loss on noisy regions is a generic OOD regularizer; combining DETA++'s training with energy-based or distance-based scoring functions at inference could compound the reported FPR95 gains, since the paper evaluates OOD detection with a single MCM head.
  • A region-level query denoiser, applying the same CoRA weights to query-image regions before LocalNCC, would directly extend the framework to noisy queries, whereas the paper currently relies on clean memory-bank centroids.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes DETA++, a plug-in module for task-adaptation-based few-shot learning (FSL) that aims to make the adapted model robust to both in-distribution (ID) noise (e.g., cluttered backgrounds) and out-of-distribution (OOD) noise (e.g., samples from unknown classes) in both the support and query sets. The method computes region and image weights with a Contrastive Relevance Aggregation (CoRA) module, then uses a clean prototype loss and a noise entropy maximization loss for adaptation. It also stores clean regions in a memory bank, uses an Intra-class Region Swapping (IntraSwap) strategy for prototype rectification, and employs a Local Nearest Centroid Classifier (LocalNCC) for inference. The paper reports extensive experiments on Meta-Dataset for few-shot classification and on ImageNet-1K for few-shot OOD detection, showing consistent improvements over six baselines and state-of-the-art results in several settings.

Significance. If the claims hold, DETA++ addresses a practically important and underexplored problem: few-shot models deployed in the open world are likely to encounter noisy support and query samples. The paper's main strengths are its broad experimental coverage (six baselines, multiple backbone architectures, and two benchmark families), component-wise ablations, Friedman test analyses, and a promised code release. The gains are consistent across baselines and noise ratios, and the proposed components are clearly motivated. However, two issues are load-bearing for the central claim that DETA++ is a general, reliable plug-in for FSL: an undefined statistic in the core CoRA module for 1-shot classes (which directly affects Table 3), and the selection of key hyperparameters on the evaluation benchmark itself. These issues need to be resolved before the paper's central claims can be fully accepted.

major comments (3)
  1. [Section 3.3, Eq. (3); Section 4.1.1; Table 3] The in-class relevance statistic φ(z_ij) is undefined when a class has exactly one support image, because I(z_ij) = Z^(c) \ Zi is then empty. Section 4.1.1 explicitly states that Table 3 uses the standard varied-way varied-shot Meta-Dataset protocol, which samples 1-shot classes; neither Section 3.3, Algorithm 1, nor the implementation details specifies a fallback for this case. As written, DETA++ cannot be evaluated on a large fraction of standard Meta-Dataset episodes, so the Table 3 state-of-the-art results are either non-reproducible or rely on undocumented behavior. Please specify the exact handling of N_c=1 (for example, including the other regions of the same image in I(z_ij), using an image-level prior, or assigning a default weight) and confirm that the Table 3 numbers are produced with that handling.
  2. [Section 4.1.1; Tables 5-6] The hyperparameters β, ϱ, and Ω are selected by maximizing average accuracy on the same Meta-Dataset evaluation benchmark used for the headline results in Tables 1-3. For example, Table 5 chooses β=0.3 and ϱ=0.3 from the ten-dataset average, and Table 6 chooses Ω=56/112 on the same benchmark. This selection on the test data inflates the reported gains and weakens the claim that DETA++ is a general plug-in; it also makes the description of CoRA as 'parameter-free' misleading, since ϱ is a tuned threshold. Please either select hyperparameters on a held-out validation split (for example, a subset of datasets or tasks) or report the sensitivity of the main conclusions to reasonable hyperparameter choices across all datasets.
  3. [Section 3.4, Eqs. (10)-(11)] If all support images of a class have accumulated weights below ϱ, the class prototype μ_c becomes the zero vector, and the cosine similarity ζ(r_j, μ_c) in Eq. (11) is undefined. This can occur under the high OOD-noise ratios in Table 2 (for example, 70%) or when a class contains only noisy samples; no fallback is described. Please add a well-defined treatment for empty or fully-filtered classes and verify that the reported results are unaffected.
minor comments (5)
  1. [Table 2] Table 2 reports only mean accuracies across the ten datasets without confidence intervals, unlike Table 1. Adding 95% confidence intervals would make the OOD-noise robustness claims easier to assess.
  2. [Section 3.3, Eqs. (5)-(6)] The notation eφ and eψ in Eqs. (5)-(6) is easy to confuse with embeddings; consider using \tilde{φ} and \tilde{ψ} for the normalized scores.
  3. [Section 3.3; Section 5] The text calls CoRA 'parameter-free,' yet the method uses hyperparameters ϱ, β, and Ω. The limitations paragraph already acknowledges extra hyperparameters; please align the terminology.
  4. [Section 4.1.1] The paragraph that introduces the fixed 10-shot protocol for Tables 1-2 and the standard varied-way varied-shot protocol for Table 3 should explicitly state how the 1-shot case is handled in the CoRA module, given Major Comment 1.
  5. [Table 3] Several baseline methods in Table 3 are cited to [24] (for example, Finetune, ProtoNet, FoProMA); please cite the original sources for these baselines where available.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: DETA++'s claims are empirical, its losses and weights are computed from the support set, and its benchmark gains are not forced by construction.

full rationale

DETA++ does not present a first-principles derivation chain; its central claim is that the proposed CoRA weighting, clean prototype loss, noise entropy loss, IntraSwap, and LocalNCC improve few-shot classification and OOD detection under dual noises. The region weights in Eqs. (3)-(7) are computed from the support-set features and then used to form the losses in Eqs. (10)-(14), which is a method specification rather than a prediction equivalent to its inputs. The experimental gains in Tables 1-3 and 8 are empirical and are compared against external baselines, so no fitted parameter is renamed as a prediction. Hyperparameters beta, rho, and Omega are tuned via ablations on the same Meta-Dataset benchmark (Tables 5-6), which raises a test-set-selection concern but does not make the reported improvements true by construction. Self-citations to DETA [1] and other author works are contextual; the new modules are supported by component-wise ablations in Tables 4 and 9, not by the citations alone. The paper's own limitation statement concedes added hyperparameters and scalability questions. The potential undefinedness of Eq. (3) when a class has one support image is a correctness issue under the varied-shot protocol, but it is not circularity. Overall, no load-bearing step reduces to its own inputs.

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

The method is empirical and introduces several tuned hyperparameters (β, ϱ, Ω, γ). It relies on the domain assumption that within-class region relevance exceeds cross-class relevance in the embedding space of pre-trained models, and that OOD images have low relevance to all classes. No new ontological entities are introduced.

free parameters (4)
  • β (noise entropy loss coefficient) = 0.3 (classification), 1.0 (OOD detection)
    Controls the contribution of Lnoise, tuned on Meta-Dataset and ImageNet-1K benchmarks to maximize reported accuracy or OOD metrics.
  • ϱ (clean/noisy threshold) = 0.3
    Threshold for separating clean from noisy regions and images, tuned on the same benchmarks where final results are reported.
  • Ω (region crop size) = 56 for 84x84 images, 112 for 224x224 images
    Size of cropped local regions, tuned to maximize performance on the evaluation datasets.
  • γ (momentum for image weights) = 0.7
    Momentum hyperparameter in the accumulator for image weights; fixed, but still a manually chosen value.
assumptions (4)
  • domain assumption Within-class region relevance is higher than cross-class relevance in the embedding space of pre-trained models.
    This assumption underpins the CoRA weight formula in Eq. (3)-(7). If false, the region weighting would not identify object regions.
  • domain assumption OOD samples have consistently low relevance to all ID classes and can be down-weighted by averaging region weights.
    The image weight in Eq. (8) and the threshold filtering in Eq. (10) rely on this. No evidence is given that this holds across all OOD sources.
  • ad hoc to paper A single threshold ϱ works across classes, datasets, and noise levels.
    Used to define clean and noisy sets in Section 3.3 and to filter images in Eq. (10). The paper tunes ϱ on the benchmarks rather than justifying it theoretically.
  • domain assumption IntraSwap (pasting clean regions into same-class images) improves prototype rectification without harmful side effects.
    Borrowed from CutMix intuition, but the paper only shows empirical gain, not a mechanism or failure analysis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reliable Few-shot Learning under Dual Noises." pith.science (2026). https://pith.science/paper/ELBEGQI4

@misc{pith2026250616330,
  author       = {Pith},
  title        = {Pith review of: Reliable Few-shot Learning under Dual Noises},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ELBEGQI4}},
  note         = {Machine review of arXiv:2506.16330}
}
read the original abstract

Recent advances in model pre-training give rise to task adaptation-based few-shot learning (FSL), where the goal is to adapt a pre-trained task-agnostic model for capturing task-specific knowledge with a few-labeled support samples of the target task.Nevertheless, existing approaches may still fail in the open world due to the inevitable in-distribution (ID) and out-of-distribution (OOD) noise from both support and query samples of the target task. With limited support samples available, i) the adverse effect of the dual noises can be severely amplified during task adaptation, and ii) the adapted model can produce unreliable predictions on query samples in the presence of the dual noises. In this work, we propose DEnoised Task Adaptation (DETA++) for reliable FSL. DETA++ uses a Contrastive Relevance Aggregation (CoRA) module to calculate image and region weights for support samples, based on which a clean prototype loss and a noise entropy maximization loss are proposed to achieve noise-robust task adaptation. Additionally,DETA++ employs a memory bank to store and refine clean regions for each inner-task class, based on which a Local Nearest Centroid Classifier (LocalNCC) is devised to yield noise-robust predictions on query samples. Moreover, DETA++ utilizes an Intra-class Region Swapping (IntraSwap) strategy to rectify ID class prototypes during task adaptation, enhancing the model's robustness to the dual noises. Extensive experiments demonstrate the effectiveness and flexibility of DETA++.

Figures

Figures reproduced from arXiv: 2506.16330 by the authors.

Figure 1
Figure 1. Dual noises in the support (training) set and query (test) set of few-shot tasks. i) ID Noise: in-distribution (ID) samples with cluttered image backgrounds. ii) OOD Noise: out-of-distribution (OOD) samples, i.e., samples from unknown classes. designing effective and efficient full-finetuning based [18], [19], adapter-finetuning based [12], [13], [20], and prompt tuning based [21], [22], [23] task adaptation algorit… view at source ↗
Figure 2
Figure 2. Quantitative evidences that both ID and OOD noise in support [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Overview of our DETA++ framework. Firstly, the images together with a set of randomly cropped local regions of support samples are fed into a pre-trained model fθ (w/ or w/o a model-specific adapter Aα) to extract image and region representations. Secondly, a contrastive relevance aggregation (CoRA) module takes the region representations as input to determine the weight of each region, based on which we can compute… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Effectiveness of our designed IntraSwap and LocalNCC in our [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison of the losses in DETA [ [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Visualization of the cropped regions and calculated weights of the CoRA module for ten 5-way 10-shot tasks sampled from Meta-Dataset. [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

106 extracted references · 68 canonical work pages

  1. [1]

    Deta: Denoised task adaptation for few-shot learning,

    J. Zhang, L. Gao, X. Luo, H. Shen, and J. Song, “Deta: Denoised task adaptation for few-shot learning,” in ICCV, October 2023, pp. 11 541–11 551

    2023

  2. [2]

    Prototypical networks for few-shot learning,

    J. Snell, K. Swersky, and R. Zemel, “Prototypical networks for few-shot learning,” in NeurIPS, 2017, pp. 4077–4087. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 15 ImageNet ImagesRegions1.00 0.90 0.55 0.43 0.15 0.11 Other 4 clas. Omglot ImagesRegions1.00 0.95 0.45 0.51 0.20 0.31 Aircraft ImagesRegions0.89 1.00 0.22 0.34 0.21 0.23 CUB ImagesRe...

    2017

  3. [3]

    Masked autoencoders are scalable vision learners,

    K. He, X. Chen, S. Xie, Y. Li, P . Doll ´ar, and R. Girshick, “Masked autoencoders are scalable vision learners,” in CVPR, 2022, pp. 16 000–16 009

    2022

  4. [4]

    Adaptive cross-modal few-shot learning,

    C. Xing, N. Rostamzadeh, B. Oreshkin, and P . O. O Pinheiro, “Adaptive cross-modal few-shot learning,” NeurIPS, vol. 32, 2019

    2019

  5. [5]

    Flex: Unifying evalua- tion for few-shot nlp,

    J. Bragg, A. Cohan, K. Lo, and I. Beltagy, “Flex: Unifying evalua- tion for few-shot nlp,” NeurIPS, vol. 34, pp. 15 787–15 800, 2021

    2021

  6. [6]

    Model-agnostic meta-learning for fast adaptation of deep networks,

    C. Finn, P . Abbeel, and S. Levine, “Model-agnostic meta-learning for fast adaptation of deep networks,” ICML, p. 1126–1135, 2017

    2017

  7. [7]

    Reinforced attention for few-shot learning and beyond,

    J. Hong, P . Fang, W. Li, T. Zhang, C. Simon, M. Harandi, and L. Petersson, “Reinforced attention for few-shot learning and beyond,” in CVPR, 2021, pp. 913–923

    2021

  8. [8]

    Meta-detr: Image- level few-shot detection with inter-class correlation exploitation,

    G. Zhang, Z. Luo, K. Cui, S. Lu, and E. P . Xing, “Meta-detr: Image- level few-shot detection with inter-class correlation exploitation,” IEEE Transactions on Pattern Analysis and Machine Intelligence , 2022

    2022

Show all 106 references
  1. [9]

    Meta-transfer learning through hard tasks,

    Q. Sun, Y. Liu, Z. Chen, T.-S. Chua, and B. Schiele, “Meta-transfer learning through hard tasks,” IEEE Transactions on Pattern Analy- sis and Machine Intelligence, vol. 44, no. 3, pp. 1443–1456, 2020

    2020

  2. [10]

    Revisiting unsupervised meta- learning via the characteristics of few-shot tasks,

    H.-J. Ye, L. Han, and D.-C. Zhan, “Revisiting unsupervised meta- learning via the characteristics of few-shot tasks,” IEEE Transac- tions on Pattern Analysis and Machine Intelligence, vol. 45, no. 3, pp. 3721–3737, 2022

    2022

  3. [11]

    A closer look at few-shot classification,

    W.-Y. Chen, Y.-C. Liu, Z. Kira, Y.-C. F. Wang, and J.-B. Huang, “A closer look at few-shot classification,” ICLR, 2019

    2019

  4. [12]

    Universal representation learning from multiple domains for few-shot classification,

    W.-H. Li and e. a. Liu, “Universal representation learning from multiple domains for few-shot classification,” in ICCV, 2021, pp. 9526–9535

    2021

  5. [13]

    Cross-domain few-shot learning with task-specific adapters,

    W.-H. Li, X. Liu, and H. Bilen, “Cross-domain few-shot learning with task-specific adapters,” in CVPR, 2022, pp. 7161–7170

    2022

  6. [14]

    A closer look at few-shot classification again,

    X. Luo, H. Wu, J. Zhang, L. Gao, J. Xu, and J. Song, “A closer look at few-shot classification again,” in ICML. PMLR, 2023, pp. 23 103–23 123

    2023

  7. [15]

    Momentum con- trast for unsupervised visual representation learning,

    K. He, H. Fan, Y. Wu, S. Xie, and R. Girshick, “Momentum con- trast for unsupervised visual representation learning,” in CVPR, 2020, pp. 9729–9738

    2020

  8. [16]

    Transformer in transformer,

    K. Han, A. Xiao, E. Wu, J. Guo, C. Xu, and Y. Wang, “Transformer in transformer,” NeurIPS, vol. 34, pp. 15 908–15 919, 2021

    2021

  9. [17]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” in ICCV, 2021, pp. 10 012–10 022

    2021

  10. [18]

    A baseline for few-shot image classification,

    G. S. Dhillon, P . Chaudhari, A. Ravichandran, and S. Soatto, “A baseline for few-shot image classification,” in ICLR, 2020

    2020

  11. [19]

    Pushing the limits of simple pipelines for few-shot learning: External data and fine-tuning make a difference,

    S. X. Hu, D. Li, J. St ¨uhmer, M. Kim, and T. M. Hospedales, “Pushing the limits of simple pipelines for few-shot learning: External data and fine-tuning make a difference,” in CVPR, 2022, pp. 9068–9077

    2022

  12. [20]

    Exploring efficient few-shot adaptation for vision transformers,

    C. Xu, S. Yang, Y. Wang, Z. Wang, Y. Fu, and X. Xue, “Exploring efficient few-shot adaptation for vision transformers,” Transac- tions on Machine Learning Research, 2022

    2022

  13. [21]

    Prompt-aligned gradient for prompt tuning,

    B. Zhu, Y. Niu, Y. Han, Y. Wu, and H. Zhang, “Prompt-aligned gradient for prompt tuning,” ICCV, 2023

    2023

  14. [22]

    Prompt learning with optimal transport for vision-language models,

    G. Chen, W. Yao, X. Song, X. Li, Y. Rao, and K. Zhang, “Prompt learning with optimal transport for vision-language models,” ICLR, 2022

    2022

  15. [23]

    Visual prompt tuning,

    M. Jia, L. Tang, B.-C. Chen, C. Cardie, S. Belongie, B. Hariharan, and S.-N. Lim, “Visual prompt tuning,” in ECCV, 2022, pp. 709– 727

    2022

  16. [24]

    Meta- dataset: A dataset of datasets for learning to learn from few examples,

    E. Triant., T. Zhu, V . Dumoulin, P . Lamblin, U. Evci, K. Xu, R. Goroshin, C. Gelada, K. Swersky, P .-A. Manzagolet al., “Meta- dataset: A dataset of datasets for learning to learn from few examples,” ICLR, 2020

    2020

  17. [25]

    Classification with noisy labels by importance reweighting,

    T. Liu and D. Tao, “Classification with noisy labels by importance reweighting,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 38, no. 3, pp. 447–461, 2015

    2015

  18. [26]

    Learning from noisy labels with deep neural networks: A survey,

    H. Song, M. Kim, D. Park, Y. Shin, and J.-G. Lee, “Learning from noisy labels with deep neural networks: A survey,” IEEE Transactions on Neural Networks and Learning Systems, 2022

    2022

  19. [27]

    Resolving training biases via influence-based data relabeling,

    S. Kong, Y. Shen, and L. Huang, “Resolving training biases via influence-based data relabeling,” in ICLR, 2021

    2021

  20. [28]

    Locoop: Few-shot out- of-distribution detection via prompt learning,

    A. Miyai, Q. Yu, G. Irie, and K. Aizawa, “Locoop: Few-shot out- of-distribution detection via prompt learning,” NeurIPS, vol. 36, 2024

    2024

  21. [29]

    Libfewshot: A comprehensive library for few-shot learning,

    W. Li, Z. Wang, X. Yang, C. Dong, P . Tian, T. Qin, J. Huo, Y. Shi, L. Wang, Y. Gao et al. , “Libfewshot: A comprehensive library for few-shot learning,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023

    2023

  22. [30]

    Few-shot object detection and viewpoint estimation for objects in the wild,

    Y. Xiao, V . Lepetit, and R. Marlet, “Few-shot object detection and viewpoint estimation for objects in the wild,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 45, no. 3, pp. 3090–3106, 2022

    2022

  23. [31]

    How to trust unlabeled data? instance credibility inference for few-shot learning,

    Y. Wang, L. Zhang, Y. Yao, and Y. Fu, “How to trust unlabeled data? instance credibility inference for few-shot learning,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 44, no. 10, pp. 6240–6253, 2021

    2021

  24. [32]

    Knowledge-guided multi-label few-shot learning for general image recognition,

    T. Chen, L. Lin, R. Chen, X. Hui, and H. Wu, “Knowledge-guided multi-label few-shot learning for general image recognition,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 44, no. 3, pp. 1371–1384, 2020

    2020

  25. [33]

    On first-order meta- learning algorithms,

    A. Nichol, J. Achiam, and J. Schulman, “On first-order meta- learning algorithms,” arXiv preprint arXiv:1803.02999, 2018

    2018 arXiv

  26. [34]

    Meta networks,

    T. Munkhdalai and H. Yu, “Meta networks,” in ICML, 2017, pp. 2554–2563

    2017

  27. [35]

    Cross attention network for few-shot classification,

    R. Hou, H. Chang, B. Ma, S. Shan, and X. Chen, “Cross attention network for few-shot classification,” NeurIPS, vol. 32, 2019

    2019

  28. [36]

    Few-shot learning via embedding adaptation with set-to-set functions,

    H.-J. Ye, H. Hu, D.-C. Zhan, and F. Sha, “Few-shot learning via embedding adaptation with set-to-set functions,” in CVPR, 2020, pp. 8808–8817

    2020

  29. [37]

    Deepemd: Differentiable earth mover’s distance for few-shot learning,

    C. Zhang, Y. Cai, G. Lin, and C. Shen, “Deepemd: Differentiable earth mover’s distance for few-shot learning,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 45, no. 5, pp. 5632–5648, 2022

    2022

  30. [38]

    A broader study of cross- domain few-shot learning,

    Y. Guo, N. C. Codella, L. Karlinsky, J. V . Codella, J. R. Smith, K. Saenko, T. Rosing, and R. Feris, “A broader study of cross- domain few-shot learning,” in ECCV, 2020, pp. 124–141

    2020

  31. [39]

    Cross- domain few-shot classification via learned feature-wise transfor- mation,

    H.-Y. Tseng, H.-Y. Lee, J.-B. Huang, and M.-H. Yang, “Cross- domain few-shot classification via learned feature-wise transfor- mation,” ICLR, 2020. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 16

    2020

  32. [40]

    Boosting the generalization capability in cross-domain few-shot learning via noise-enhanced supervised autoencoder,

    H. Liang, Q. Zhang, P . Dai, and J. Lu, “Boosting the generalization capability in cross-domain few-shot learning via noise-enhanced supervised autoencoder,” in ICCV, 2021, pp. 9424–9434

    2021

  33. [41]

    Tgdm: Target guided dynamic mixup for cross-domain few-shot learning,

    L. Zhuo, Y. Fu, J. Chen, Y. Cao, and Y.-G. Jiang, “Tgdm: Target guided dynamic mixup for cross-domain few-shot learning,” in Proceedings of the 30th ACM International Conference on Multimedia, 2022, pp. 6368–6376

    2022

  34. [42]

    Generalized meta-fdmixup: Cross-domain few-shot learning guided by labeled target data,

    Y. Fu, Y. Fu, J. Chen, and Y.-G. Jiang, “Generalized meta-fdmixup: Cross-domain few-shot learning guided by labeled target data,” IEEE Transactions on Image Processing, vol. 31, pp. 7078–7090, 2022

    2022

  35. [43]

    Free-lunch for cross- domain few-shot learning: Style-aware episodic training with robust contrastive learning,

    J. Zhang, J. Song, L. Gao, and H. Shen, “Free-lunch for cross- domain few-shot learning: Style-aware episodic training with robust contrastive learning,” in ACM MM, 2022, pp. 2586–2594

    2022

  36. [44]

    Styleadv: Meta style ad- versarial training for cross-domain few-shot learning,

    Y. Fu, Y. Xie, Y. Fu, and Y.-G. Jiang, “Styleadv: Meta style ad- versarial training for cross-domain few-shot learning,” in CVPR, 2023, pp. 24 575–24 584

    2023

  37. [45]

    Fast and flexible multi-task classification using conditional neu- ral adaptive processes,

    J. Requeima, J. Gordon, J. Bronskill, S. Nowozin, and R. E. Turner, “Fast and flexible multi-task classification using conditional neu- ral adaptive processes,” NeurIPS, vol. 32, 2019

    2019

  38. [46]

    Improved few-shot visual classification,

    P . Bateni, R. Goyal, V . Masrani, F. Wood, and L. Sigal, “Improved few-shot visual classification,” in CVPR, 2020, pp. 14 493–14 502

    2020

  39. [47]

    Learning a universal template for few-shot dataset generaliza- tion,

    E. Triantafillou, H. Larochelle, R. Zemel, and V . Dumoulin, “Learning a universal template for few-shot dataset generaliza- tion,” in ICML, 2021, pp. 10 424–10 433

    2021

  40. [48]

    Dense classifi- cation and implanting for few-shot learning,

    Y. Lifchitz, Y. Avrithis, S. Picard, and A. Bursuc, “Dense classifi- cation and implanting for few-shot learning,” in CVPR, 2019, pp. 9258–9267

    2019

  41. [49]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agar- wal, G. Sastry, A. Askell, P . Mishkin, J. Clark et al. , “Learning transferable visual models from natural language supervision,” in ICML, 2021, pp. 8748–8763

    2021

  42. [50]

    Learning to prompt for vision-language models,

    K. Zhou, J. Yang, C. C. Loy, and Z. Liu, “Learning to prompt for vision-language models,” International Journal of Computer Vision, vol. 130, no. 9, pp. 2337–2348, 2022

    2022

  43. [51]

    Conditional prompt learning for vision-language models,

    K. Zhou, J. Yang, and et. al, “Conditional prompt learning for vision-language models,” in CVPR, 2022, pp. 16 816–16 825

    2022

  44. [52]

    Visual-language prompt tuning with knowledge-guided context optimization,

    H. Yao, R. Zhang, and C. Xu, “Visual-language prompt tuning with knowledge-guided context optimization,” in CVPR, 2023, pp. 6757–6767

    2023

  45. [53]

    Dept: Decoupled prompt tuning,

    J. Zhang, S. Wu, L. Gao, H. Shen, and J. Song, “Dept: Decoupled prompt tuning,” CVPR, 2024

    2024

  46. [54]

    A closer look at memorization in deep networks,

    D. Arpit, S. Jastrzebski, N. Ballas, D. Krueger, E. Bengio, M. S. Kanwal, T. Maharaj, A. Fischer, A. Courville, Y. Bengio et al., “A closer look at memorization in deep networks,” ICML, pp. 233– 242, 2017

    2017

  47. [55]

    Rectifying the shortcut learning of background for few-shot learning,

    X. Luo, L. Wei, L. Wen, J. Yang, L. Xie, Z. Xu, and Q. Tian, “Rectifying the shortcut learning of background for few-shot learning,” NeurIPS, vol. 34, pp. 13 073–13 085, 2021

    2021

  48. [56]

    Few-shot learning with noisy labels,

    K. J. Liang, S. B. Rangrej, V . Petrovic, and T. Hassner, “Few-shot learning with noisy labels,” in CVPR, 2022, pp. 9089–9098

    2022

  49. [57]

    Deepemd: Few-shot image classification with differentiable earth mover’s distance and structured classifiers,

    C. Zhang, Y. Cai, G. Lin, and C. Shen, “Deepemd: Few-shot image classification with differentiable earth mover’s distance and structured classifiers,” in CVPR, 2020, pp. 12 203–12 213

    2020

  50. [58]

    Toward open set recognition,

    W. J. Scheirer, A. de Rezende Rocha, A. Sapkota, and T. E. Boult, “Toward open set recognition,”IEEE transactions on pattern analysis and machine intelligence, vol. 35, no. 7, pp. 1757–1772, 2012

    2012

  51. [59]

    Towards open world recognition,

    A. Bendale and T. Boult, “Towards open world recognition,” in CVPR, 2015, pp. 1893–1902

    2015

  52. [60]

    Towards open set deep networks,

    A. Bendale and T. E. Boult, “Towards open set deep networks,” in CVPR, 2016, pp. 1563–1572

    2016

  53. [61]

    Enhancing the reliability of out- of-distribution image detection in neural networks,

    S. Liang, Y. Li, and R. Srikant, “Enhancing the reliability of out- of-distribution image detection in neural networks,” ICLR, 2018

    2018

  54. [62]

    Energy-based out-of- distribution detection,

    W. Liu, X. Wang, J. Owens, and Y. Li, “Energy-based out-of- distribution detection,” NeurIPS, vol. 33, pp. 21 464–21 475, 2020

    2020

  55. [63]

    A baseline for detecting misclassi- fied and out-of-distribution examples in neural networks,

    D. Hendrycks and K. Gimpel, “A baseline for detecting misclassi- fied and out-of-distribution examples in neural networks,” ICLR, 2017

    2017

  56. [64]

    Mitigating neural network overconfidence with logit normalization,

    H. Wei, R. Xie, H. Cheng, L. Feng, B. An, and Y. Li, “Mitigating neural network overconfidence with logit normalization,” in ICML, 2022, pp. 23 631–23 644

    2022

  57. [65]

    Dice: Leveraging sparsification for out-of- distribution detection,

    Y. Sun and Y. Li, “Dice: Leveraging sparsification for out-of- distribution detection,” in ECCV, 2022

    2022

  58. [66]

    Line: Out-of-distribution detection by leveraging important neurons,

    Y. H. Ahn, G.-M. Park, and S. T. Kim, “Line: Out-of-distribution detection by leveraging important neurons,” in CVPR, 2023, pp. 19 852–19 862

    2023

  59. [67]

    From global to local: Multi-scale out-of-distribution detection,

    J. Zhang, L. Gao, B. Hao, H. Huang, J. Song, and H. Shen, “From global to local: Multi-scale out-of-distribution detection,” IEEE Transactions on Image Processing, 2023

    2023

  60. [68]

    Poodle: Improving few-shot learning via penalizing out-of-distribution samples,

    D. Le, K. D. Nguyen, K. Nguyen, Q.-H. Tran, R. Nguyen, and B.-S. Hua, “Poodle: Improving few-shot learning via penalizing out-of-distribution samples,” in NeurIPS, M. Ranzato, A. Beygelz- imer, Y. Dauphin, P . Liang, and J. W. Vaughan, Eds., vol. 34, 2021, pp. 23 942–23 955

    2021

  61. [69]

    Distance- based image classification: Generalizing to new classes at near- zero cost,

    T. Mensink, J. Verbeek, F. Perronnin, and G. Csurka, “Distance- based image classification: Generalizing to new classes at near- zero cost,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 35, no. 11, pp. 2624–2637, 2013

    2013

  62. [70]

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

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly et al., “An image is worth 16x16 words: Transformers for image recognition at scale,” ICLR, 2020

    2020

  63. [71]

    Do better imagenet models transfer better?

    S. Kornblith, J. Shlens, and Q. V . Le, “Do better imagenet models transfer better?” in CVPR, 2019, pp. 2661–2671

    2019

  64. [72]

    A simple framework for contrastive learning of visual representations,

    T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A simple framework for contrastive learning of visual representations,” in ICML. PMLR, 2020, pp. 1597–1607

    2020

  65. [73]

    Learning contrastive embedding in low-dimensional space,

    S. Chen, C. Gong, J. Li, J. Yang, G. Niu, and M. Sugiyama, “Learning contrastive embedding in low-dimensional space,” NeurIPS, vol. 35, pp. 6345–6357, 2022

    2022

  66. [74]

    Meta-learning with latent embedding optimization,

    A. A. Rusu, D. Rao, J. Sygnowski, O. Vinyals, R. Pascanu, S. Osin- dero, and R. Hadsell, “Meta-learning with latent embedding optimization,” ICLR, 2019

    2019

  67. [75]

    Prototype rectification for few-shot learning,

    J. Liu, L. Song, and Y. Qin, “Prototype rectification for few-shot learning,” in ECCV. Springer, 2020, pp. 741–756

    2020

  68. [76]

    Alleviating the sample selection bias in few-shot learning by removing projection to the centroid,

    J. Xu, X. Luo, X. Pan, W. Pei, Y. Li, and Z. Xu, “Alleviating the sample selection bias in few-shot learning by removing projection to the centroid,” NeurIPS, 2022

    2022

  69. [77]

    Cutmix: Regularization strategy to train strong classifiers with localizable features,

    S. Yun, D. Han, S. J. Oh, S. Chun, J. Choe, and Y. Yoo, “Cutmix: Regularization strategy to train strong classifiers with localizable features,” in ICCV, 2019, pp. 6023–6032

    2019

  70. [78]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” in CVPR. Ieee, 2009, pp. 248–255

    2009

  71. [79]

    A universal representation transformer layer for few-shot image classification,

    L. Liu, W. Hamilton, G. Long, J. Jiang, and H. Larochelle, “A universal representation transformer layer for few-shot image classification,” ICLR, 2021

    2021

  72. [80]

    Emerging properties in self-supervised vision transformers,

    M. Caron, H. Touvron, I. Misra, H. J ´egou, J. Mairal, P . Bojanowski, and A. Joulin, “Emerging properties in self-supervised vision transformers,” ICCV, pp. 9650–9660, 2021

    2021

  73. [81]

    Training data-efficient image transformers-distillation through attention,

    H. Touvron, M. Cord, M. Douze, F. Massa, A. Sablayrolles, and H. Jegou, “Training data-efficient image transformers-distillation through attention,” in ICML, 2021, pp. 10 347–10 357

    2021

  74. [83]

    Optimized generic fea- ture learning for few-shot classification across domains,

    T. Saikia, T. Brox, and C. Schmid, “Optimized generic fea- ture learning for few-shot classification across domains,” arXiv preprint arXiv:2001.07926, 2020

    2001 arXiv

  75. [84]

    Enhanc- ing few-shot image classification with unlabelled examples,

    P . Bateni, J. Barber, J.-W. van de Meent, and F. Wood, “Enhanc- ing few-shot image classification with unlabelled examples,” in WACVW, 2022, pp. 2796–2805

    2022

  76. [85]

    Selecting relevant features from a multi-domain representation for few-shot classification,

    N. Dvornik, C. Schmid, and J. Mairal, “Selecting relevant features from a multi-domain representation for few-shot classification,” in ECCV, 2020, pp. 769–786

    2020

  77. [86]

    A multi- mode modulator for multi-domain few-shot classification,

    Y. Liu, J. Lee, L. Zhu, L. Chen, H. Shi, and Y. Yang, “A multi- mode modulator for multi-domain few-shot classification,” in ICCV, 2021, pp. 8453–8462

    2021

  78. [87]

    Supervised contrastive learning,

    P . Khosla, P . Teterwak, C. Wang, A. Sarna, Y. Tian, P . Isola, A. Maschinot, C. Liu, and D. Krishnan, “Supervised contrastive learning,” NeurIPS, vol. 33, pp. 18 661–18 673, 2020

    2020

  79. [88]

    The use of ranks to avoid the assumption of normality implicit in the analysis of variance,

    M. Friedman, “The use of ranks to avoid the assumption of normality implicit in the analysis of variance,” Journal of the american statistical association, vol. 32, no. 200, pp. 675–701, 1937

    1937

  80. [89]

    Advanced nonparametric tests for multiple comparisons in the design of experiments in computational intelligence and data mining: Experimental analysis of power,

    S. Garc ´ıa, A. Fern ´andez, J. Luengo, and F. Herrera, “Advanced nonparametric tests for multiple comparisons in the design of experiments in computational intelligence and data mining: Experimental analysis of power,” Information sciences , vol. 180, no. 10, pp. 2044–2064, 2010

    2010

  81. [90]

    Skip tuning: Pre-trained vision-language models are effective and efficient adapters themselves,

    S. Wu, J. Zhang, P . Zeng, L. Gao, J. Song, and H. T. Shen, “Skip tuning: Pre-trained vision-language models are effective and efficient adapters themselves,” CVPR, 2025. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 17

    2025

  82. [91]

    Delving into out- of-distribution detection with vision-language representations,

    Y. Ming, Z. Cai, J. Gu, Y. Sun, W. Li, and Y. Li, “Delving into out- of-distribution detection with vision-language representations,” NeurIPS, vol. 35, pp. 35 087–35 102, 2022

    2022

  83. [92]

    Scaling for training time and post-hoc out-of-distribution detection enhancement,

    K. Xu, R. Chen, G. Franchi, and A. Yao, “Scaling for training time and post-hoc out-of-distribution detection enhancement,” arXiv preprint arXiv:2310.00227, 2023

    2023 arXiv

  84. [93]

    Zero-shot in-distribution detection in multi-object settings using vision-language founda- tion models,

    A. Miyai, Q. Yu, G. Irie, and K. Aizawa, “Zero-shot in-distribution detection in multi-object settings using vision-language founda- tion models,” arXiv preprint arXiv:2304.04521, 2023

    2023 arXiv

  85. [94]

    Nearest neighbor guidance for out-of-distribution detection,

    J. Park, Y. G. Jung, and A. B. J. Teoh, “Nearest neighbor guidance for out-of-distribution detection,” in Proceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 1686–1695

    2023

  86. [95]

    Enhancing the reliability of out-of- distribution image detection in neural networks,

    S. Liang, Y. Li, and R. Srikant, “Enhancing the reliability of out-of- distribution image detection in neural networks,” in ICLR, 2018

    2018

  87. [96]

    Vim: Out-of-distribution with virtual-logit matching,

    H. Wang, Z. Li, L. Feng, and W. Zhang, “Vim: Out-of-distribution with virtual-logit matching,” in CVPR, 2022, pp. 4921–4930

    2022

  88. [97]

    Out-of-distribution detection with deep nearest neighbors,

    Y. Sun, Y. Ming, X. Zhu, and Y. Li, “Out-of-distribution detection with deep nearest neighbors,” in ICML. PMLR, 2022, pp. 20 827– 20 840

    2022

  89. [98]

    Non-parametric outlier synthe- sis,

    L. Tao, X. Du, J. Zhu, and Y. Li, “Non-parametric outlier synthe- sis,” in ICLR, 2022

    2022

  90. [99]

    Mos: Towards scaling out-of-distribution detection for large semantic space,

    R. Huang and Y. Li, “Mos: Towards scaling out-of-distribution detection for large semantic space,” in CVPR, 2021, pp. 8710– 8719

    2021

  91. [100]

    The inaturalist species classification and detection dataset,

    G. Van Horn, O. Mac Aodha, Y. Song, Y. Cui, C. Sun, A. Shepard, H. Adam, P . Perona, and S. Belongie, “The inaturalist species classification and detection dataset,” in CVPR, 2018, pp. 8769– 8778

    2018

  92. [101]

    Sun database: Large-scale scene recognition from abbey to zoo,

    J. Xiao, J. Hays, K. A. Ehinger, A. Oliva, and A. Torralba, “Sun database: Large-scale scene recognition from abbey to zoo,” in 2010 IEEE computer society conference on computer vision and pattern recognition. IEEE, 2010, pp. 3485–3492

    2010

  93. [102]

    Places: A 10 million image database for scene recognition,

    B. Zhou, A. Lapedriza, A. Khosla, A. Oliva, and A. Torralba, “Places: A 10 million image database for scene recognition,”IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 40, no. 6, pp. 1452–1464, 2017

    2017

  94. [103]

    Describing textures in the wild,

    M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, and A. Vedaldi, “Describing textures in the wild,” in CVPR, 2014, pp. 3606–3613

    2014

  95. [104]

    In or out? fixing im- agenet out-of-distribution detection evaluation,

    J. Bitterwolf, M. Mueller, and M. Hein, “In or out? fixing im- agenet out-of-distribution detection evaluation,” arXiv preprint arXiv:2306.00826, 2023

    2023 arXiv

  96. [105]

    Entropic out-of-distribution detection: Seamless detection of unknown examples,

    D. Mac ˆedo, T. I. Ren, C. Zanchettin, A. L. Oliveira, and T. Luder- mir, “Entropic out-of-distribution detection: Seamless detection of unknown examples,” IEEE Transactions on Neural Networks and Learning Systems, vol. 33, no. 6, pp. 2350–2364, 2021

    2021

  97. [106]

    Open-set recog- nition: a good closed-set classifier is all you need?

    S. Vaze, K. Han, A. Vedaldi, and A. Zisserman, “Open-set recog- nition: a good closed-set classifier is all you need?” in ICLR, 2022

    2022

  98. [107]

    Exploring the limits of out-of-distribution detection,

    S. Fort, J. Ren, and B. Lakshminarayanan, “Exploring the limits of out-of-distribution detection,” NeurIPS, vol. 34, pp. 7068–7081, 2021. Ji Zhang is an Assistant Professor with the School of Computing and Artificial Intelligence, Southwest Jiaotong University, China. He ob- t...

    2021

Pith tools

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