pith. sign in

arxiv: 2605.25110 · v1 · pith:5XIMMDEOnew · submitted 2026-05-24 · 💻 cs.CV · cs.AI· cs.LG

Uncertainty-DTW for Sequences and Visual Tokens

Lei Wang , Syuan-Hao Li , Yongsheng Gao , Piotr Koniusz This is my paper

Pith reviewed 2026-06-30 11:56 UTC · model grok-4.3

classification 💻 cs.CV cs.AIcs.LG
keywords uncertainty-DTWdynamic time warpingvisual tokensprobabilistic alignmentsequence alignmentnoise robustnessattention mechanisms
0
0 comments X

The pith

Uncertainty-DTW models each match as a Normal distribution to suppress unreliable features during sequence and visual token alignment.

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

The paper introduces uncertainty-aware alignment as a probabilistic framework for matching structured data such as time series and visual representations. It formulates uncertainty-DTW (uDTW) by assigning a Normal distribution to each pairwise correspondence and optimizing alignment paths through a maximum likelihood objective. The objective uses a precision-weighted matching term to downweight noisy or unreliable features along with log-variance regularization to avoid degenerate solutions. The same mechanism extends from temporal sequences to sets of visual tokens, where the resulting uncertainty values act as a reverse-attention signal that highlights semantically relevant regions. Experiments across domains report consistent gains over prior alignment methods and show that the learned uncertainties track semantic importance.

Core claim

Our formulation, uncertainty-DTW (uDTW), assigns each correspondence a Normal distribution and parametrizes each alignment path by a Maximum Likelihood Estimate objective consisting of (i) a precision-weighted matching term that suppresses unreliable features, and (ii) a log-variance regularization that prevents degenerate solutions. This yields a probabilistic alignment mechanism that is robust to noise and interpretable, as uncertainty directly reflects the reliability of matches. We further generalize this framework from temporal sequences to tokenized visual representations.

What carries the argument

uncertainty-DTW (uDTW), which assigns independent Normal distributions to pairwise correspondences and optimizes each alignment path via an MLE objective combining precision-weighted matching with log-variance regularization

If this is right

  • Alignments become robust to heterogeneous and noisy features through the precision-weighted term.
  • Uncertainty values directly indicate match reliability and serve as an interpretable reverse-attention signal.
  • Structured matching extends naturally from sequences to unordered sets of visual tokens.
  • Evaluations show that learned uncertainty correlates with semantic importance across domains.

Where Pith is reading between the lines

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

  • The reverse-attention property could let uncertainty scores replace hand-crafted attention maps in downstream vision pipelines.
  • The same Normal-distribution formulation might stabilize differentiable alignment layers inside larger neural networks.
  • If the variance regularization works as claimed, similar uncertainty terms could be added to other matching objectives such as optimal transport.

Load-bearing premise

Modeling each pairwise correspondence as an independent Normal distribution with learnable variance, combined with the proposed MLE objective, produces alignments that are robust to noise and semantically meaningful.

What would settle it

An experiment on noisy sequence or visual-token datasets in which uDTW alignments show no improvement over deterministic DTW or in which learned uncertainties fail to correlate with unreliable or ambiguous regions would falsify the central claim.

Figures

Figures reproduced from arXiv: 2605.25110 by Lei Wang, Piotr Koniusz, Syuan-Hao Li, Yongsheng Gao.

Figure 1
Figure 1. Figure 1: A connection between alignment, attention, and uncertainty: ( [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: (a)-(d) Alignment paths of sDTW and uDTW (white) for a pair of sequences. Pixel intensities are power-normalized (exponent 0.1) to enhance [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of visual-token embeddings before and after the learned uDTW projection. Token embeddings are visualized by mapping the first [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Uncertainty-aware alignment across tasks. (a) Time series forecasting: predicted sequences are aligned with ground truth using uDTW with learned [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Interpolation between two time series (grey and black dashed) on the [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of barycenters computed with sDTW and uDTW on CBF [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison between DAAM attention and the learned uncertainty across MiniImageNet, Stanford Cars, CUB-200-2011, UFGVC, and Aircraft. [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
read the original abstract

Aligning structured data is a fundamental problem in computer vision and machine learning, underlying tasks such as time series analysis, human action recognition, and visual representation learning. Existing alignment methods, including Dynamic Time Warping (DTW) and its differentiable variants, rely on deterministic similarity measures and are therefore sensitive to heterogeneous and noisy features. In this work, we introduce uncertainty-aware alignment, a probabilistic framework that models pairwise correspondences with heteroscedastic uncertainty and performs structured matching along alignment paths. Our formulation, uncertainty-DTW (uDTW), assigns each correspondence a Normal distribution and parametrizes each alignment path by a Maximum Likelihood Estimate objective consisting of (i) a precision-weighted matching term that suppresses unreliable features, and (ii) a log-variance regularization that prevents degenerate solutions. This yields a probabilistic alignment mechanism that is robust to noise and interpretable, as uncertainty directly reflects the reliability of matches. We further generalize this framework from temporal sequences to tokenized visual representations, enabling structured matching over sets of visual tokens. The learned uncertainty can be interpreted as a reverse-attention: semantically relevant regions exhibit low uncertainty and dominate the alignment, while ambiguous/noisy regions have high uncertainty. This provides a connection between alignment, attention, and uncertainty modeling. We evaluate the proposed framework across diverse domains. The results demonstrate consistent improvements over state-of-the-art methods and show that learned uncertainty correlates with semantic importance. These findings establish uncertainty-aware alignment as a general, robust, and interpretable framework for learning from structured data.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 2 minor

Summary. The manuscript introduces uncertainty-DTW (uDTW), a probabilistic framework for sequence and visual token alignment. It models each pairwise correspondence as a Normal distribution with learnable heteroscedastic variance and selects paths via an MLE objective consisting of a precision-weighted matching term (to suppress unreliable features) plus a log-variance regularizer (to avoid degeneracies). The approach is extended to tokenized visual representations, with learned uncertainty interpreted as reverse-attention that highlights semantically relevant regions. Experiments across domains are reported to show consistent gains over SOTA methods and correlation between uncertainty and semantic importance.

Significance. If the central formulation holds, the work offers a principled probabilistic treatment of uncertainty within DTW-style alignment, potentially improving robustness in noisy CV tasks such as action recognition and token matching. The explicit link between alignment, uncertainty, and attention mechanisms is a conceptual contribution that could influence future hybrid models, provided the probabilistic justification and empirical gains are substantiated.

major comments (2)
  1. [§3 (MLE objective)] The MLE objective (abstract and §3 formulation) factorizes the log-likelihood as a sum of independent per-correspondence terms (precision-weighted squared error + log-variance). Because valid alignment paths are constrained by monotonicity and continuity (standard DTW recurrence), the selected matches are statistically dependent; the independence assumption therefore renders the joint likelihood misspecified. This directly affects the claimed suppression of unreliable features and the reverse-attention interpretation, as the regularization term cannot compensate for the missing path-level joint structure.
  2. [§3 and experimental sections] No derivation or justification is supplied for why the constrained-path likelihood can be safely factorized, nor is there an ablation comparing the proposed objective against a properly joint formulation (e.g., via dynamic programming that respects the constraints inside the likelihood). Without this, the robustness and interpretability claims rest on an unverified modeling choice.
minor comments (2)
  1. The abstract states that results demonstrate improvements and semantic correlation, yet the provided text supplies no quantitative tables, dataset names, or metric values; these must be added for reproducibility.
  2. [§3] Notation for the Normal parameters (mean, variance) and the precise form of the precision-weighted term should be introduced with explicit equations rather than prose descriptions.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive feedback on our manuscript. The comments highlight important aspects of the probabilistic formulation that warrant further clarification. We provide point-by-point responses below and will update the manuscript to address these issues.

read point-by-point responses

  1. Referee: [§3 (MLE objective)] The MLE objective (abstract and §3 formulation) factorizes the log-likelihood as a sum of independent per-correspondence terms (precision-weighted squared error + log-variance). Because valid alignment paths are constrained by monotonicity and continuity (standard DTW recurrence), the selected matches are statistically dependent; the independence assumption therefore renders the joint likelihood misspecified. This directly affects the claimed suppression of unreliable features and the reverse-attention interpretation, as the regularization term cannot compensate for the missing path-level joint structure.

    Authors: We agree that the path constraints introduce dependencies among the selected correspondences. Our model assumes that, conditional on the alignment path, the correspondences are independent, with each pair having its own heteroscedastic variance. The dynamic programming procedure then selects the path that maximizes the factorized likelihood. This factorization enables efficient computation and is analogous to the emission probabilities in a hidden Markov model, where observations are independent given the state sequence, but the sequence is constrained by transition probabilities. We will revise §3 to include a formal derivation of the likelihood under this conditional independence assumption and discuss its implications for the uncertainty interpretation. The reverse-attention view follows from the learned variances reflecting match reliability within the selected path. revision: yes

  2. Referee: [§3 and experimental sections] No derivation or justification is supplied for why the constrained-path likelihood can be safely factorized, nor is there an ablation comparing the proposed objective against a properly joint formulation (e.g., via dynamic programming that respects the constraints inside the likelihood). Without this, the robustness and interpretability claims rest on an unverified modeling choice.

    Authors: As noted above, we will add the derivation and justification in the revised manuscript. For the ablation, a fully joint formulation would involve computing the marginal likelihood over all valid paths, which is computationally prohibitive for long sequences. Our approach uses the maximum a posteriori path under the factorized model, which is standard in DTW variants. We will include a discussion of this approximation and, if feasible within the revision timeline, an ablation on short sequences comparing to a brute-force joint computation. This will substantiate the modeling choice. revision: partial

Circularity Check

0 steps flagged

No circularity: uDTW is a direct modeling proposal using standard MLE on independent Normals

full rationale

The abstract and description present uDTW as a new formulation that assigns Normal distributions to pairwise correspondences and defines an MLE objective consisting of a precision-weighted matching term plus log-variance regularization. This is a modeling choice grounded in standard probabilistic assumptions rather than any reduction of a claimed result to fitted inputs or self-citations. No equations, derivations, or load-bearing steps are shown that equate outputs to inputs by construction. The independence assumption and path parametrization are explicit design decisions, not hidden circularities. The paper is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on the domain assumption that correspondences are well-modeled by independent Normal distributions whose variances can be learned to reflect reliability. No free parameters or invented entities are identifiable from the abstract alone.

axioms (2)
  • domain assumption Pairwise correspondences follow independent Normal distributions with heteroscedastic variances.
    Explicitly stated in the abstract as the modeling choice for each correspondence.
  • domain assumption Alignment paths can be parametrized by maximizing the likelihood under the stated MLE objective.
    The abstract defines the alignment path via this MLE objective consisting of the two terms.

pith-pipeline@v0.9.1-grok · 5805 in / 1416 out tokens · 44917 ms · 2026-06-30T11:56:41.250461+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

81 extracted references · 3 canonical work pages · 3 internal anchors

  1. [1]

    Dynamic accumulated attention map for interpreting evolution of decision-making in vision transformer,

    Y . Liao, Y . Gao, and W. Zhang, “Dynamic accumulated attention map for interpreting evolution of decision-making in vision transformer,”Pattern Recognition, vol. 165, p. 111607, 2025

    2025

  2. [2]

    Fast global alignment kernels,

    M. Cuturi, “Fast global alignment kernels,” inInternational Conference on Machine Learning (ICML), 2011, pp. 929–936

    2011

  3. [3]

    Soft-dtw: a differentiable loss function for time-series,

    M. Cuturi and M. Blondel, “Soft-dtw: a differentiable loss function for time-series,” inInternational Conference on Machine Learning (ICML). PMLR, 2017, pp. 894–903

    2017

  4. [4]

    Temporal-viewpoint transportation plan for skeletal few-shot action recognition,

    L. Wang and P. Koniusz, “Temporal-viewpoint transportation plan for skeletal few-shot action recognition,” inAsian conference on computer vision (ACCV), 2022, pp. 4176–4193

    2022

  5. [5]

    Uncertainty-dtw for time series and sequences,

    ——, “Uncertainty-dtw for time series and sequences,” inEuropean Conference on Computer Vision (ECCV), 2022, pp. 176–195

    2022

  6. [6]

    Meet jeanie: a similarity measure for 3d skeleton sequences via temporal-viewpoint alignment,

    L. Wang, J. Liu, L. Zheng, T. Gedeon, and P. Koniusz, “Meet jeanie: a similarity measure for 3d skeleton sequences via temporal-viewpoint alignment,”International Journal of Computer Vision, vol. 132, no. 9, pp. 4091–4122, 2024

    2024

  7. [7]

    Differentiable divergences be- tween time series,

    M. Blondel, A. Mensch, and J.-P. Vert, “Differentiable divergences be- tween time series,” inInternational Conference on Artificial Intelligence and Statistics (AISTATS), 2021, pp. 3853–3861

    2021

  8. [8]

    Drop- dtw: Aligning common signal between sequences while dropping out- liers,

    M. Dvornik, I. Hadji, K. G. Derpanis, A. Garg, and A. Jepson, “Drop- dtw: Aligning common signal between sequences while dropping out- liers,”Advances in Neural Information Processing Systems (NeurIPS), vol. 34, pp. 13 782–13 793, 2021

    2021

  9. [9]

    Temporal alignment prediction for supervised representation learning and few-shot sequence classification,

    B. Su and J.-R. Wen, “Temporal alignment prediction for supervised representation learning and few-shot sequence classification,” inInter- national Conference on Learning Representations (ICLR), 2022

    2022

  10. [10]

    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. Gellyet al., “An image is worth 16x16 words: Transformers for image recognition at scale,” inInternational Conference on Learning Representations (ICLR), 2020

    2020

  11. [11]

    Dynam- icvit: Efficient vision transformers with dynamic token sparsification,

    Y . Rao, W. Zhao, B. Liu, J. Lu, J. Zhou, and C.-J. Hsieh, “Dynam- icvit: Efficient vision transformers with dynamic token sparsification,” Advances in Neural Information Processing Systems (NeurIPS), vol. 34, pp. 13 937–13 949, 2021

    2021

  12. [12]

    Motion meets attention: Video motion prompts,

    Q. Chen, L. Wang, P. Koniusz, and T. Gedeon, “Motion meets attention: Video motion prompts,” inAsian Conference on Machine Learning (ACML), 2024, pp. 591–606

    2024

  13. [13]

    Dtwnet: A dynamic time warping network,

    X. Cai, T. Xu, J. Yi, J. Huang, and S. Rajasekaran, “Dtwnet: A dynamic time warping network,”Advances in Neural Information Processing Systems (NeurIPS), vol. 32, 2019

    2019

  14. [14]

    Dtw-nn: A novel neural network for time series recognition using dynamic alignment between inputs and weights,

    B. K. Iwana, V . Frinken, and S. Uchida, “Dtw-nn: A novel neural network for time series recognition using dynamic alignment between inputs and weights,”Knowledge-Based Systems, vol. 188, p. 104971, 2020. 12

    2020

  15. [15]

    Deep attentive time warping,

    S. Matsuo, X. Wu, G. Atarsaikhan, A. Kimura, K. Kashino, B. K. Iwana, and S. Uchida, “Deep attentive time warping,”Pattern Recognition, vol. 136, p. 109201, 2023

    2023

  16. [16]

    Recent advances in optimal transport for machine learning,

    E. F. Montesuma, F. M. N. Mboula, and A. Souloumiac, “Recent advances in optimal transport for machine learning,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 47, no. 2, pp. 1161– 1180, 2024

    2024

  17. [17]

    Sinkhorn distances: Lightspeed computation of opti- mal transport,

    M. Cuturi, “Sinkhorn distances: Lightspeed computation of opti- mal transport,”Advances in Neural Information Processing Systems (NeurIPS), vol. 26, 2013

    2013

  18. [18]

    Hierarchical optimal transport for multimodal distribution alignment,

    J. Lee, M. Dabagia, E. Dyer, and C. Rozell, “Hierarchical optimal transport for multimodal distribution alignment,”Advances in Neural Information Processing Systems (NeurIPS), vol. 32, 2019

    2019

  19. [19]

    Spatio-temporal alignments: Optimal transport through space and time,

    H. Janati, M. Cuturi, and A. Gramfort, “Spatio-temporal alignments: Optimal transport through space and time,” inInternational Conference on Artificial Intelligence and Statistics (AISTATS). PMLR, 2020, pp. 1695–1704

    2020

  20. [20]

    Toward accurate dynamic time warping in linear time and space,

    S. Salvador and P. Chan, “Toward accurate dynamic time warping in linear time and space,”Intelligent Data Analysis, vol. 11, no. 5, pp. 561–580, 2007

    2007

  21. [21]

    Dynamic time warping algorithm review,

    P. Senin, “Dynamic time warping algorithm review,”Information and Computer Science Department University of Hawaii at Manoa Honolulu, USA, vol. 855, no. 1-23, p. 40, 2008

    2008

  22. [22]

    Time warp edit distance with stiffness adjustment for time series matching,

    P.-F. Marteau, “Time warp edit distance with stiffness adjustment for time series matching,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 31, no. 2, pp. 306–318, 2008

    2008

  23. [23]

    Deep metric learning using triplet network,

    E. Hoffer and N. Ailon, “Deep metric learning using triplet network,” inInternational Workshop on Similarity-Based Pattern Recognition. Springer, 2015, pp. 84–92

    2015

  24. [24]

    Superglue: Learning feature matching with graph neural networks,

    P.-E. Sarlin, D. DeTone, T. Malisiewicz, and A. Rabinovich, “Superglue: Learning feature matching with graph neural networks,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 4938–4947

    2020

  25. [25]

    Token-label alignment for vision transformers,

    H. Xiao, W. Zheng, Z. Zhu, J. Zhou, and J. Lu, “Token-label alignment for vision transformers,” inProceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 5495–5504

    2023

  26. [26]

    Scale-aware token-matching for transformer-based object detector,

    A. Jung, S. Hong, and Y . Hyun, “Scale-aware token-matching for transformer-based object detector,”Pattern Recognition Letters, vol. 185, pp. 197–202, 2024

    2024

  27. [27]

    Madtp: Multi- modal alignment-guided dynamic token pruning for accelerating vision- language transformer,

    J. Cao, P. Ye, S. Li, C. Yu, Y . Tang, J. Lu, and T. Chen, “Madtp: Multi- modal alignment-guided dynamic token pruning for accelerating vision- language transformer,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 15 710–15 719

    2024

  28. [28]

    Simple spectral graph convolution,

    H. Zhu and P. Koniusz, “Simple spectral graph convolution,” in International Conference on Learning Representations (ICLR), 2021. [Online]. Available: https://openreview.net/forum?id=CYO5T-YjWZV

    2021

  29. [29]

    In defense of soft-assignment coding,

    L. Liu, L. Wang, and X. Liu, “In defense of soft-assignment coding,” in The IEEE International Conference on Computer Vision (ICCV), 2011, pp. 2486–2493

    2011

  30. [30]

    The ucr time series archive,

    H. A. Dau, A. Bagnall, K. Kamgar, C.-C. M. Yeh, Y . Zhu, S. Gharghabi, C. A. Ratanamahatana, and E. Keogh, “The ucr time series archive,” IEEE/CAA Journal of Automatica Sinica, vol. 6, no. 6, pp. 1293–1305, 2019

    2019

  31. [31]

    Ntu rgb+ d: A large scale dataset for 3d human activity analysis,

    A. Shahroudy, J. Liu, T.-T. Ng, and G. Wang, “Ntu rgb+ d: A large scale dataset for 3d human activity analysis,” inThe IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 1010– 1019

    2016

  32. [32]

    Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding,

    J. Liu, A. Shahroudy, M. Perez, G. Wang, L.-Y . Duan, and A. C. Kot, “Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 42, no. 10, pp. 2684–2701, 2019

    2019

  33. [33]

    The Kinetics Human Action Video Dataset

    W. Kay, J. Carreira, K. Simonyan, B. Zhang, C. Hillier, S. Vijaya- narasimhan, F. Viola, T. Green, T. Back, P. Natsevet al., “The kinetics human action video dataset,”arXiv preprint arXiv:1705.06950, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  34. [34]

    Matching networks for one shot learning,

    O. Vinyals, C. Blundell, T. Lillicrap, D. Wierstraet al., “Matching networks for one shot learning,”Advances in Neural Information Pro- cessing Systems (NeurIPS), vol. 29, 2016

    2016

  35. [35]

    Prototypical networks for few- shot learning,

    J. Snell, K. Swersky, and R. Zemel, “Prototypical networks for few- shot learning,”Advances in Neural Information Processing Systems (NeurIPS), vol. 30, 2017

    2017

  36. [36]

    Meta-learning with differentiable closed-form solvers,

    L. Bertinetto, J. F. Henriques, P. Torr, and A. Vedaldi, “Meta-learning with differentiable closed-form solvers,” inInternational Conference on Learning Representations (ICLR), 2019. [Online]. Available: https://openreview.net/forum?id=HyxnZh0ct7

    2019

  37. [37]

    Meta-learning for semi-supervised few-shot classification,

    M. Ren, S. Ravi, E. Triantafillou, J. Snell, K. Swersky, J. B. Tenenbaum, H. Larochelle, and R. S. Zemel, “Meta-learning for semi-supervised few-shot classification,” inInternational Conference on Learning Representations (ICLR), 2018. [Online]. Available: https://openreview.net/forum?id=HJcSzz-CZ

    2018

  38. [38]

    The caltech-ucsd birds-200-2011 dataset,

    C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie, “The caltech-ucsd birds-200-2011 dataset,” 2011

    2011

  39. [39]

    Novel dataset for fine-grained image categorization: Stanford dogs,

    A. Khosla, N. Jayadevaprakash, B. Yao, and F.-F. Li, “Novel dataset for fine-grained image categorization: Stanford dogs,” inThe IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), vol. 2, no. 1, 2011

    2011

  40. [40]

    3d object representations for fine-grained categorization,

    J. Krause, M. Stark, J. Deng, and L. Fei-Fei, “3d object representations for fine-grained categorization,” inThe IEEE International Conference on Computer Vision Workshops (ICCVW), 2013, pp. 554–561

    2013

  41. [41]

    Fine-Grained Visual Classification of Aircraft

    S. Maji, E. Rahtu, J. Kannala, M. Blaschko, and A. Vedaldi, “Fine- grained visual classification of aircraft,”arXiv preprint arXiv:1306.5151, 2013

    work page internal anchor Pith review Pith/arXiv arXiv 2013

  42. [42]

    Building a bird recognition app and large scale dataset with citizen scientists: The fine print in fine-grained dataset collection,

    G. Van Horn, S. Branson, R. Farrell, S. Haber, J. Barry, P. Ipeirotis, P. Perona, and S. Belongie, “Building a bird recognition app and large scale dataset with citizen scientists: The fine print in fine-grained dataset collection,” inThe IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015, pp. 595–604

    2015

  43. [43]

    Benchmark platform for ultra-fine-grained visual categorization beyond human performance,

    X. Yu, Y . Zhao, Y . Gao, X. Yuan, and S. Xiong, “Benchmark platform for ultra-fine-grained visual categorization beyond human performance,” in The IEEE International Conference on Computer Vision (ICCV), 2021, pp. 10 285–10 295

    2021

  44. [44]

    Human-level concept learning through probabilistic program induction,

    B. M. Lake, R. Salakhutdinov, and J. B. Tenenbaum, “Human-level concept learning through probabilistic program induction,”Science, vol. 350, no. 6266, pp. 1332–1338, 2015

    2015

  45. [45]

    Boosting learning efficiency in few-shot tasks with layer-adaptive pid control,

    P. Zhang, X. Li, L. Yu, Z. Zhang, F. Dunkin, H. Liu, and Z. Li, “Boosting learning efficiency in few-shot tasks with layer-adaptive pid control,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2026

    2026

  46. [46]

    Class-aware patch embedding adaptation for few-shot image classification,

    F. Hao, F. He, L. Liu, F. Wu, D. Tao, and J. Cheng, “Class-aware patch embedding adaptation for few-shot image classification,” inThe IEEE International Conference on Computer Vision (ICCV), 2023, pp. 18 905– 18 915

    2023

  47. [47]

    Adaptive saliency based contextual metric learning for few-shot open-set recognition,

    P. Li, J. Chen, L. Shang, and C. Ping, “Adaptive saliency based contextual metric learning for few-shot open-set recognition,”Pattern Recognition, p. 113096, 2026

    2026

  48. [48]

    Unsupervised learning via meta- learning,

    K. Hsu, S. Levine, and C. Finn, “Unsupervised learning via meta- learning,” inInternational Conference on Learning Representations (ICLR), 2019. [Online]. Available: https://openreview.net/forum?id= r1My6sR9tX

    2019

  49. [49]

    Unsupervised meta-learning for few-shot image classification,

    S. Khodadadeh, L. Boloni, and M. Shah, “Unsupervised meta-learning for few-shot image classification,”Advances in Neural Information Processing Systems (NeurIPS), vol. 32, 2019

    2019

  50. [50]

    Unsupervised meta-learning through latent-space interpolation in generative models,

    S. Khodadadeh, S. Zehtabian, S. Vahidian, W. Wang, B. Lin, and L. Boloni, “Unsupervised meta-learning through latent-space interpolation in generative models,” inInternational Conference on Learning Representations (ICLR), 2021. [Online]. Available: https://openreview.net/forum?id=XOjv2HxIF6i

    2021

  51. [51]

    Unsupervised meta-learning via few-shot pseudo-supervised contrastive learning,

    H. Jang, H. Lee, and J. Shin, “Unsupervised meta-learning via few-shot pseudo-supervised contrastive learning,” inInternational Conference on Learning Representations (ICLR), 2023. [Online]. Available: https://openreview.net/forum?id=TdTGGj7fYYJ

    2023

  52. [52]

    Meta-GMV AE: Mixture of gaussian V AE for unsupervised meta-learning,

    D. B. Lee, D. Min, S. Lee, and S. J. Hwang, “Meta-GMV AE: Mixture of gaussian V AE for unsupervised meta-learning,” inInternational Conference on Learning Representations (ICLR), 2021. [Online]. Available: https://openreview.net/forum?id=wS0UFjsNYjn

    2021

  53. [53]

    Unsupervised meta-learning via latent space energy-based model of symbol vector coupling,

    D. Kong, B. Pang, and Y . N. Wu, “Unsupervised meta-learning via latent space energy-based model of symbol vector coupling,” inFifth Workshop on Meta-Learning at the Conference on Neural Information Processing Systems, 2021. [Online]. Available: https: //openreview.net/forum?id=-pLftu7EpXz

    2021

  54. [54]

    DINOv3

    O. Sim ´eoni, H. V . V o, M. Seitzer, F. Baldassarre, M. Oquab, C. Jose, V . Khalidov, M. Szafraniec, S. Yi, M. Ramamonjisoaet al., “Dinov3,” arXiv preprint arXiv:2508.10104, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  55. [55]

    On the limited memory bfgs method for large scale optimization,

    D. C. Liu and J. Nocedal, “On the limited memory bfgs method for large scale optimization,”Mathematical Programming, vol. 45, p. 503–528, 1989

    1989

  56. [56]

    Learning to learn task transformations for improved few-shot classification,

    G. Zheng, Q. Suo, M. Huai, and A. Zhang, “Learning to learn task transformations for improved few-shot classification,” inProceedings of the 2023 SIAM International Conference on Data Mining (SDM). SIAM, 2023, pp. 784–792

    2023

  57. [57]

    Ssl-protonet: Self- supervised learning prototypical networks for few-shot learning,

    J. Y . Lim, K. M. Lim, C. P. Lee, and Y . X. Tan, “Ssl-protonet: Self- supervised learning prototypical networks for few-shot learning,”Expert Systems with Applications, vol. 238, p. 122173, 2024

    2024

  58. [58]

    Simple semantic-aided few-shot learning,

    H. Zhang, J. Xu, S. Jiang, and Z. He, “Simple semantic-aided few-shot learning,” inThe IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 28 588–28 597. 13

    2024

  59. [59]

    Core: Correlation- guided feature enhancement for few-shot image classification,

    J. Xu, X. Pan, J. Wang, W. Pei, Q. Liao, and Z. Xu, “Core: Correlation- guided feature enhancement for few-shot image classification,”IEEE Transactions on Neural Networks and Learning Systems, vol. 36, no. 2, pp. 3098–3110, 2024

    2024

  60. [60]

    Anrot-helanet: adverserially and naturally robust attention-based ag- gregation network via the hellinger distance for few-shot classification,

    G. Y . Lee, T. Dam, M. M. Ferdaus, D. P. Poenar, and V . N. Duong, “Anrot-helanet: adverserially and naturally robust attention-based ag- gregation network via the hellinger distance for few-shot classification,” International Journal of Multimedia Information Retrieval, vol. 15, no. 1, p. 8, 2026

    2026

  61. [61]

    Few-shot classification with fork attention adapter,

    J. Sun and J. Li, “Few-shot classification with fork attention adapter,” Pattern Recognition, vol. 156, p. 110805, 2024

    2024

  62. [62]

    Tnpnet: An approach to few- shot open-set recognition via contextual transductive learning,

    S. Wu, H. Luo, and X. Lin, “Tnpnet: An approach to few- shot open-set recognition via contextual transductive learning,” Neurocomputing, vol. 621, p. 129276, 2025. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0925231224020472

    2025

  63. [63]

    Hela-vfa: A hellinger distance-attention-based feature aggregation net- work for few-shot classification,

    G. Y . Lee, T. Dam, D. P. Poenar, V . N. Duong, and M. M. Ferdaus, “Hela-vfa: A hellinger distance-attention-based feature aggregation net- work for few-shot classification,” inThe IEEE Winter Conference on Applications of Computer Vision (WACV), 2024, pp. 2173–2183

    2024

  64. [64]

    Metadiff: Meta-learning with conditional diffusion for few-shot learning,

    B. Zhang, C. Luo, D. Yu, X. Li, H. Lin, Y . Ye, and B. Zhang, “Metadiff: Meta-learning with conditional diffusion for few-shot learning,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, 2024, pp. 16 687–16 695

    2024

  65. [65]

    Adaptive feature selection-based feature reconstruction network for few-shot learning,

    J. Ren, Y . An, T. Lei, J. Yang, W. Zhang, Z. Pan, Y . Liao, Y . Gao, C. Sun, and W. Zhang, “Adaptive feature selection-based feature reconstruction network for few-shot learning,”Pattern Recognition, p. 112289, 2025

    2025

  66. [66]

    Boosting few-shot fine-grained recognition with background suppression and foreground alignment,

    Z. Zha, H. Tang, Y . Sun, and J. Tang, “Boosting few-shot fine-grained recognition with background suppression and foreground alignment,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 33, no. 8, pp. 3947–3961, 2023

    2023

  67. [67]

    Multi-level correlation network for few-shot image classification,

    Y . Dang, M. Sun, M. Zhang, Z. Chen, X. Zhang, Z. Wang, and D. Wang, “Multi-level correlation network for few-shot image classification,” in 2023 IEEE International Conference on Multimedia and Expo (ICME). IEEE, 2023, pp. 2909–2914

    2023

  68. [68]

    Efficient group attentive learn- ing for few-shot image classification,

    J. Sun, K. Huang, D. Yang, and H. Liu, “Efficient group attentive learn- ing for few-shot image classification,”Expert Systems with Applications, p. 131245, 2026

    2026

  69. [69]

    Charting the right manifold: Manifold mixup for few-shot learning,

    P. Mangla, N. Kumari, A. Sinha, M. Singh, B. Krishnamurthy, and V . N. Balasubramanian, “Charting the right manifold: Manifold mixup for few-shot learning,” inThe IEEE Winter Conference on Applications of Computer Vision (WACV), 2020, pp. 2218–2227

    2020

  70. [70]

    Robust saliency-aware dis- tillation for few-shot fine-grained visual recognition,

    H. Liu, C. P. Chen, X. Gong, and T. Zhang, “Robust saliency-aware dis- tillation for few-shot fine-grained visual recognition,”IEEE Transactions on Multimedia, vol. 26, pp. 7529–7542, 2024

    2024

  71. [71]

    T2l: Trans-transfer learning for few-shot fine- grained visual categorization with extended adaptation,

    N. Sun and P. Yang, “T2l: Trans-transfer learning for few-shot fine- grained visual categorization with extended adaptation,”Knowledge- Based Systems, vol. 264, p. 110329, 2023

    2023

  72. [72]

    Tripletmaml: A metric-based model-agnostic meta-learning algorithm for few-shot classification,

    A. G ¨ulc¨u, Z. Kus ¸, ˙I. T. S. ¨Ozkan, and O. F. Karakus ¸, “Tripletmaml: A metric-based model-agnostic meta-learning algorithm for few-shot classification,”Progress in Artificial Intelligence, pp. 1–15, 2026

    2026

  73. [73]

    Fine-grained few shot learning with foreground object transformation,

    C. Wang, S. Song, Q. Yang, X. Li, and G. Huang, “Fine-grained few shot learning with foreground object transformation,”Neurocomputing, vol. 466, pp. 16–26, 2021

    2021

  74. [74]

    Bi-directional task-guided network for few-shot fine-grained image classification,

    Z.-X. Ma, Z.-D. Chen, L.-J. Zhao, Z.-C. Zhang, T. Zheng, X. Luo, and X.-S. Xu, “Bi-directional task-guided network for few-shot fine-grained image classification,” inProceedings of the 32nd ACM International Conference on Multimedia, 2024, pp. 8277–8286

    2024

  75. [75]

    Bsnet: Bi- similarity network for few-shot fine-grained image classification,

    X. Li, J. Wu, Z. Sun, Z. Ma, J. Cao, and J.-H. Xue, “Bsnet: Bi- similarity network for few-shot fine-grained image classification,”IEEE Transactions on Image Processing, vol. 30, pp. 1318–1331, 2020

    2020

  76. [76]

    Few-shot classification with feature map reconstruction networks,

    D. Wertheimer, L. Tang, and B. Hariharan, “Few-shot classification with feature map reconstruction networks,” inThe IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021, pp. 8012– 8021

    2021

  77. [77]

    Crosstransformers: spatially- aware few-shot transfer,

    C. Doersch, A. Gupta, and A. Zisserman, “Crosstransformers: spatially- aware few-shot transfer,”Advances in Neural Information Processing Systems (NeurIPS), vol. 33, pp. 21 981–21 993, 2020

    2020

  78. [78]

    Interpretable few-shot image classification via prototypical concept-guided mixture of lora experts,

    Z. Ji, R. Wei, J. Liu, Y . Pang, and J. Han, “Interpretable few-shot image classification via prototypical concept-guided mixture of lora experts,” IEEE Transactions on Image Processing, 2026

    2026

  79. [79]

    Distribution consistency based covariance metric networks for few-shot learning,

    W. Li, J. Xu, J. Huo, L. Wang, Y . Gao, and J. Luo, “Distribution consistency based covariance metric networks for few-shot learning,” in Proceedings of the AAAI conference on Artificial Intelligence, vol. 33, 2019, pp. 8642–8649

    2019

  80. [80]

    Low-rank pairwise alignment bilinear network for few-shot fine-grained image classifica- tion,

    H. Huang, J. Zhang, J. Zhang, J. Xu, and Q. Wu, “Low-rank pairwise alignment bilinear network for few-shot fine-grained image classifica- tion,”IEEE Transactions on Multimedia, vol. 23, pp. 1666–1680, 2020

    2020

Showing first 80 references.