Pith. sign in

REVIEW 5 major objections 4 minor 67 references

Deep Convolutional Large-Margin $\ell_p$-SVDD for Visual Anomaly Detection

T0 review · 5 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read The paper argues that jointly learning a convolutional representation and an explicit large-margin ℓ_p-SVDD boundary, through alternating Frank-Wolfe boundary updates and a smooth margin-violation CNN loss, consistently outperforms fixed-fe

desk verdict Joint deep feature learning with explicit ℓp-SVDD boundary is plausible, but the claimed consistent gains over the [9] baseline are not controlled; the frozen-CNN ablation is the real evidence. read the letter →

arxiv 2607.22212 v1 pith:6GPYWKJH submitted 2026-07-24 cs.CV cs.LG

Alireza Dastmalchi Saei , Shervin Rahimzadeh Arashloo This is my paper · ORCID
classification cs.CVcs.LG
keywords anomalydetectionlarge-marginSVDDℓ_pnormdeepCNNalternatingoptimizationFrank-Wolfekernelapproximationlong-tailedrecognition
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 proposes DLM-SVDD, a method that trains a convolutional network and a large-margin ℓ_p-SVDD decision boundary together rather than separately. It alternates between solving the convex dual boundary problem with Frank-Wolfe and updating the network with a softplus margin-violation loss, so the representation adapts to the boundary and the boundary is re-fit to the new features. The authors claim this joint scheme consistently improves over the fixed-feature large-margin ℓ_p-SVDD baseline on CIFAR-10 and CIFAR-100 at every tested anomaly ratio, and that it performs strongly on long-tailed imbalance benchmarks. The practical point is that representation adaptation and explicit margin geometry can be combined in one stable training loop, and kernel approximations allow the method to scale.

What carries the argument

The alternating two-block optimization is the core mechanism: an α-step that solves the convex ℓ_p-SVDD dual via Frank-Wolfe with a closed-form linear subproblem, and a ω-step that backpropagates a smooth margin-violation loss through the CNN. The margin-violation gradient focuses updates on samples that cross the inner or outer boundary, and low-rank kernel backends (Nyström, random Fourier features, RPCholesky) make the kernel operations scalable.

What would settle it

Run the alternating scheme with the paper's fixed Adam learning rate and record squared parameter differences across epochs; if ||ω^(t+1) − ω^(t)|| does not tend to zero even as the loss decreases, then a premise of the stationarity proof is violated. A simpler check: if the CNN gradient norm plateaus above zero while the Frank-Wolfe gap goes to zero, the network block is not reaching the claimed stationarity.

Watch

Extended reading notes

Core claim

DLM-SVDD retains the exact large-margin ℓ_p-SVDD dual objective for the boundary while making the features trainable. At each epoch, features from the CNN define a kernel matrix, and a Frank-Wolfe solver recovers the optimal dual variables; then, with that boundary frozen, the CNN is updated to push normal samples inside the inner boundary and anomalous samples outside the outer boundary, using a softplus-regularized version of the slack penalties. The paper reports that this joint training lifts AUROC by up to about 0.117 over a frozen backbone and beats the fixed-feature baseline at all tested ratios, with the biggest gains under severe class imbalance.

Load-bearing premise

The convergence guarantee rests on the assumption that consecutive CNN parameter updates shrink to zero during training; the paper assumes this 'asymptotic regularity' rather than proving it, and Adam with a fixed learning rate makes it nontrivial.

Editorial extensions

If this is right

  • If the claims hold, anomaly detectors can get both adaptive representations and explicit margin geometry without sacrificing either.
  • Joint training gives the largest gains under severe class imbalance, suggesting representation adaptation matters most when the normal class is poorly represented.
  • Kernel approximations such as Nyström and RPCholesky recover near-exact accuracy at a fraction of memory and time, making the approach practical on large datasets.
  • The alternating recipe of boundary re-fitting followed by margin-aware feature updates is a template that could extend to other kernel-based one-class objectives.

Reading between the lines

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

  • The paper's own ImageNet-LT results hint that joint training can hurt when positive samples are extremely scarce; a hybrid that freezes the backbone for very small classes might combine the best of both.
  • Fixing the RBF bandwidth once from pretrained features avoids a feedback loop; a testable extension would be to adapt the bandwidth slowly or with a separate objective to get further gains.
  • Because the CNN loss only has gradient support near the boundary (via the softplus sigmoid), the method may be naturally insensitive to noisy labels far from the boundary—an untested but plausible side effect.
  • The convergence proof assumes successive CNN updates shrink to zero; in practice, using a decaying learning rate schedule would bring the algorithm in line with the theory.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 4 minor

Summary. The paper proposes DLM-SVDD, a deep anomaly-detection framework that alternates between solving a large-margin ℓ_p-SVDD dual boundary (via Frank–Wolfe on current CNN features) and updating the CNN backbone with a smooth margin-violation loss. It reports experiments on standard one-vs-rest benchmarks, long-tailed CIFAR/ImageNet variants, and a systematic study of seven kernel approximation backends. The central claim is that joint representation–boundary learning consistently improves over the fixed-feature ℓ_p-SVDD baseline [9] and achieves the best average rank among compared methods.

Significance. If the central claim holds, the paper makes a useful contribution: it connects an explicit large-margin kernel boundary to deep representation learning in a single alternating scheme, and it provides a practical scalability analysis of kernel approximations within that scheme. Strengths include the public code link, the controlled kernel-approximation study (Fig. 3, Table IV), the use of the Skillings–Mack rank test, and the internal fixed-CNN ablation (Table V). However, the main empirical claim is currently undercut by an uncontrolled comparison against the [9] baseline and by missing variance reporting; the optimization claim is also weakened by an unproven regularity assumption.

major comments (5)
  1. [Tables I, II, and V] The headline claim 'consistently improves over the fixed-feature ℓ_p-SVDD baseline of [9]' is not a controlled comparison. Table I reports [9] at 0.958 AUROC on CIFAR-10 (ratio 10%), while the paper's own fixed-CNN DLM-SVDD variant, using the same ImageNet-pretrained ResNet-50 and training only the boundary, reaches 0.867 (Table V). Table II shows an even larger gap: 89.20 vs 59.50 balanced accuracy on CIFAR-10-LT (ϱ=100). Both rows are described as fixed-feature large-margin ℓ_p-SVDD. Unless [9] used a different feature extractor or protocol, these numbers are inconsistent. Rerun [9] with the same pretrained features, or restrict the claim to the within-method fixed-vs-joint improvement.
  2. [Section III-C2 and Section IV-B] Eq. (7) writes L_ω as a sum of ψ^p terms, but Section IV-B states that p=1 is used in the softplus for all experiments, and Eq. (8) gives the gradient as σ(v), not pψ^{p-1}σ(v). Thus the CNN step actually minimizes a p=1 softplus margin loss while the boundary solver uses p=2. This is a mismatch between the stated ℓ_p objective and the implemented loss. Please correct Eqs. (7)–(8), state the p=1 choice explicitly, and justify it; the current text describes the method as optimizing the ℓ_p primal for general p>1 in the CNN update.
  3. [Section III-F] The convergence statement (17) depends on 'assume asymptotic regularity, ||ω^{(t+1)}−ω^{(t)}||→0', which is neither proved nor implied by A1–A2 and the descent inequality (16). Adam with constant learning rate does not guarantee this, and the boundary re-estimation changes the objective each epoch. Please either prove this condition under explicit step-size/decay assumptions or present (17) clearly as a conditional statement. Figure 2 shows loss curves for two classes and does not verify the regularity condition.
  4. [Tables I and II] Results are reported as ten-seed means without standard deviations or confidence intervals. For example, the CIFAR-10 ratio-75 gap between DLM-SVDD (0.984) and Single kernel [9] (0.972) is 0.012 AUROC; without variance, the 'consistent improvement' and rank claims are hard to evaluate. Report per-cell standard deviations/confidence intervals and, where possible, pairwise significance tests. The Skillings–Mack omnibus p-value only establishes that some rank difference exists.
  5. [Section IV-B] The common configuration p=2, ν=1.2 is selected on the validation split and then used for all reported test results. This can bias the comparison if baselines were not tuned to the same extent. Please specify the validation-based selection rule, report sensitivity of final AUROC to p and ν at least at one ratio, and state the tuning budget for each compared method.
minor comments (4)
  1. [Figure 2] Please specify which CIFAR classes and seeds are shown and add variability across runs; otherwise 'representative behavior' is unverifiable.
  2. [Table IV caption] The caption says 'single seed per class', so the ±std is across the 10 classes, not across seeds. State this explicitly to avoid misinterpretation.
  3. [Section IV-D / Table II] ImageNet-LT results appear not to be averaged over ten seeds. Clarify the seed protocol and report variance for those rows.
  4. [Eq. (14)] For RFF, the standard construction includes a normalization factor and often a paired sine term to obtain unbiased kernel estimates. Clarify whether the displayed map is the one used in all spectral backends and whether any normalization/scaling is applied.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the boundary dual is a cited building block from the authors' prior work, but the central joint-learning claim is an empirical comparison, not a definitional reduction.

full rationale

The paper's derivation chain is not circular in the sense prohibited by the rubric. The boundary update (Eq. 2) is the dual of the large-margin ℓp-SVDD objective taken from the authors' prior work [9], and the Frank–Wolfe solver is applied to that published, peer-reviewed formulation. This is a normal reliance on a cited theorem, not a self-referential reduction. The CNN update (Eq. 7) is a softplus surrogate of the slack terms explicitly inherited from the primal objective (Eq. 1/6), so the representation loss is constructed from the boundary constraints rather than from the claimed result. The central claim that joint representation–boundary learning improves over a fixed backbone is tested in Table V, where both 'fixed CNN' and 'joint' variants share the same pretrained ResNet-50 and pipeline; the positive ΔAUROC is an empirical outcome, not an identity forced by a fitted constant or by definition. No parameter is fitted to a subset and then reported as a prediction of a closely related quantity. The convergence analysis (Section III-F) does contain an explicit unproven assumption—'assume asymptotic regularity, ||ω^{(t+1)}−ω^{(t)}||→0'—and omits a proof of that condition; however, this is a proof gap/weakness, not circularity. Likewise, the baseline comparison to [9] in Tables I–II may be confounded (e.g., 'DLM-SVDD (fixed CNN)' differs markedly from 'Single-kernel ℓp-SVDD [9]' in Table II), and [9]/[16] are self-citations, but these are experimental-validity concerns rather than circular reductions. Because no equation in the paper reduces to its own input by construction, the circularity score is low.

Assumptions & free parameters 5 free parameters · 7 assumptions · 0 invented entities

The central method rests on the prior dual of [9], standard kernel-approximation theory, and a smooth-surrogate assumption. No new physical or mathematical entities are introduced. The main ad hoc elements are the unproven asymptotic-regularity condition and the p=1 softplus surrogate.

free parameters (5)
  • RBF bandwidth γ = median of pairwise squared distances of pretrained features
    Set once from pretrained training features and held fixed during alternating training; controls kernel geometry and the boundary shape.
  • margin parameter ν = 1.2
    Selected based on validation observations and fixed as common configuration; appears in the feasible set constraint 1^T α = ν.
  • slack-norm exponent p = p=2 for boundary; p=1 for softplus CNN loss
    p=2 is used in the boundary update; the CNN update uses softplus with p=1 after preliminary experiments, changing the effective penalty for violations.
  • learning rate η = 1e-4
    Adam learning rate fixed as common configuration across experiments.
  • kernel approximation budget m = varies 64–4096
    In approximation experiments m is a chosen budget; for main results the exact kernel is presumably used, but the ImageNet-LT backend is not specified.
assumptions (7)
  • domain assumption The dual in Eq. (2) is the exact dual of the primal large-margin ℓp-SVDD problem in Eq. (1).
    Section III-C.1 relies on the dual derivation from [9] without re-deriving it; this is a cited prior result, not proven in this paper.
  • domain assumption Assumptions A1 (bounded iterates) and A2 (Lipschitz smoothness of f_ω) hold during training.
    Section III-F states these standard conditions; A1 is said to be 'promoted' by weight decay and gradient clipping, not proven.
  • ad hoc to paper The alternating iterates satisfy asymptotic regularity ||ω^{(t+1)}−ω^{(t)}||→0.
    Assumed for the alternating convergence result in Section III-F; no proof is provided, and with a constant learning rate Adam this is not guaranteed in general.
  • standard math Frank–Wolfe convergence for convex objectives with a linear subproblem solver.
    Uses Jaggi's Frank–Wolfe convergence theorem to bound the boundary-solver error.
  • standard math RBF kernel is shift-invariant, positive definite, and satisfies k(z,z)=1.
    Required for the score formula f_ω(x)=1−2k(z)^T a + a^T K a in Eq. (10).
  • standard math Kernel approximation methods (Nyström, RFF, QMC-RFF, ORF/SORF, Fastfood, RPCholesky) provide valid low-rank or spectral approximations to the RBF kernel.
    Section III-E relies on standard Bochner and Nyström theory without re-deriving approximation guarantees.
  • domain assumption Softplus is an adequate smooth surrogate for the hinge loss in the CNN update.
    Section III-C.2 replaces the powered hinge with softplus and then sets p=1 in practice; the representation update therefore optimizes a surrogate, not the exact ℓp slack loss.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deep Convolutional Large-Margin $\ell_p$-SVDD for Visual Anomaly Detection." pith.science (2026). https://pith.science/paper/6GPYWKJH

@misc{pith2026260722212,
  author       = {Pith},
  title        = {Pith review of: Deep Convolutional Large-Margin $\ell_p$-SVDD for Visual Anomaly Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6GPYWKJH}},
  note         = {Machine review of arXiv:2607.22212}
}
abstract

Visual anomaly detection requires adaptive representations and reliable decision boundaries, particularly when anomalous training samples are scarce and class distributions are highly imbalanced. Classical kernel-based methods yield principled geometric decision regions but typically operate on fixed features, while deep detectors learn task-specific representations but often fail to provide an explicit margin-aware kernel boundary. In this study, we propose DLM-SVDD, a deep large-margin novelty-detection framework that jointly learns convolutional features and an explicit kernel-based decision boundary. By drawing on the large-margin $\ell_p$-Support Vector Data Description ($\ell_p$-SVDD) approach, the proposed method performs explicit margin maximization and nonlinear slack penalization while adapting the representation to the target task. To train the proposed model, we present an optimization scheme that alternates between a Frank--Wolfe--based update of the convex dual boundary and a CNN update step operating on a smooth margin-violation loss induced by the recovered boundary. To improve scalability, we analyze the efficiency--accuracy trade-offs for different kernel approximation strategies, deriving practical propositions for large-scale anomaly detection. Extensive experiments on multiple standard benchmarks show consistent performance improvements over the baseline and strong overall performance compared with state-of-the-art methods while illustrating that the proposed joint representation--boundary learning scheme remains effective under severe imbalanced class distributions.

Figures

Figures reproduced from arXiv: 2607.22212 by the authors.

Figure 1
Figure 1. Overview of the proposed DLM-SVDD alternating optimization framework. At each epoch, the CNN backbone extracts deep features from all [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Representative blockwise optimization behavior for DLM-SVDD on [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Kernel approximation accuracy–efficiency analysis on CIFAR-10 (negative-to-positive ratio = 50%; see also Table IV for selected operating points). [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

67 extracted references · 2 linked inside Pith

  1. [9]

    Large-margin multiple kernelℓ p-svdd using frank–wolfe algorithm for novelty detection,

    S. Rahimzadeh Arashloo, “Large-margin multiple kernelℓ p-svdd using frank–wolfe algorithm for novelty detection,”Pattern Recognition, vol. 148, p. 110189, 2024

    2024

  2. [1]

    Panda: Adapting pretrained features for anomaly detection and segmentation,

    T. Reiss, N. Cohen, L. Bergman, and Y . Hoshen, “Panda: Adapting pretrained features for anomaly detection and segmentation,” inPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021, pp. 2806–2814

    2021

  3. [2]

    Mean-shifted contrastive loss for anomaly de- tection,

    T. Reiss and Y . Hoshen, “Mean-shifted contrastive loss for anomaly de- tection,” inProceedings of the AAAI Conference on Artificial Intelligence (AAAI), vol. 37, no. 2, 2023, pp. 2155–2162

    2023

  4. [3]

    Robust one-class clas- sification using deep kernel spectral regression,

    S. Mohammad and S. Rahimzadeh Arashloo, “Robust one-class clas- sification using deep kernel spectral regression,”Neurocomputing, vol. 573, p. 127246, 2024

    2024

  5. [4]

    Support vector data description,

    D. M. J. Tax and R. P. W. Duin, “Support vector data description,” Machine Learning, vol. 54, no. 1, pp. 45–66, 2004

    2004

  6. [5]

    Deep one-class classification,

    L. Ruff, R. A. Vandermeulen, N. G ¨ornitz, L. Deecke, S. A. Siddiqui, A. Binder, E. M ¨uller, and M. Kloft, “Deep one-class classification,” in Proceedings of the 35th International Conference on Machine Learning (ICML), ser. Proceedings of Machine Learning Research, vol. 80, 2018, pp. 4393–4402

    2018

  7. [6]

    Deep semi-supervised anomaly detection,

    L. Ruff, R. A. Vandermeulen, N. G ¨ornitz, A. Binder, E. M ¨uller, K.-R. M¨uller, and M. Kloft, “Deep semi-supervised anomaly detection,” in International Conference on Learning Representations (ICLR), 2020

    2020

  8. [7]

    Dasvdd: Deep autoencoding support vector data descriptor for anomaly detection,

    H. Hojjati and N. Armanfard, “Dasvdd: Deep autoencoding support vector data descriptor for anomaly detection,”IEEE Transactions on Knowledge and Data Engineering, 2023

    2023

Show all 67 references
  1. [8]

    Deep multi-sphere support vector data description based on disentangled representation learning,

    H.-J. Xing, H.-N. Wu, and P.-P. Zhang, “Deep multi-sphere support vector data description based on disentangled representation learning,” Pattern Recognition, vol. 156, p. 110842, 2024

    2024

  2. [10]

    Using the nystr ¨om method to speed up kernel machines,

    C. K. I. Williams and M. Seeger, “Using the nystr ¨om method to speed up kernel machines,” inAdvances in Neural Information Processing Systems 13, T. K. Leen, T. G. Dietterich, and V . Tresp, Eds. MIT Press, 2001, pp. 682–688

    2001

  3. [11]

    Random features for large-scale kernel machines,

    A. Rahimi and B. Recht, “Random features for large-scale kernel machines,” inAdvances in Neural Information Processing Systems, J. Platt, D. Koller, Y . Singer, and S. Roweis, Eds., vol. 20. Curran Associates, Inc., 2007

    2007

  4. [12]

    Quasi-monte carlo feature maps for shift-invariant kernels,

    J. Yang, V . Sindhwani, H. Avron, and M. Mahoney, “Quasi-monte carlo feature maps for shift-invariant kernels,” inProceedings of the 31st International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, E. P. Xing and T. Jebara, Eds., vol. 32, no. 1...

    2014

  5. [13]

    Orthogonal random features,

    F. X. X. Yu, A. T. Suresh, K. M. Choromanski, D. N. Holtmann-Rice, and S. Kumar, “Orthogonal random features,” inAdvances in Neural Information Processing Systems, vol. 29, 2016

    2016

  6. [14]

    Fastfood - computing hilbert space expansions in loglinear time,

    Q. Le, T. Sarlos, and A. Smola, “Fastfood - computing hilbert space expansions in loglinear time,” inProceedings of the 30th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, S. Dasgupta and D. McAllester, Eds., vol. 28, no. 3. Atlant...

    2013

  7. [15]

    Randomly pivoted cholesky: Practical approximation of a kernel matrix with few entry evaluations,

    Y . Chen, E. N. Epperly, J. A. Tropp, and R. J. Webber, “Randomly pivoted cholesky: Practical approximation of a kernel matrix with few entry evaluations,”Communications on Pure and Applied Mathematics, vol. 78, no. 5, pp. 995–1041, 2025

    2025

  8. [16]

    Deep large-margin lp-svdd with cnn feature learning for novelty detection,

    A. D. Saei and S. Rahimzadeh Arashloo, “Deep large-margin lp-svdd with cnn feature learning for novelty detection,” inICASSP 2026 - 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2026, pp. 2376–2380

    2026

  9. [17]

    Learning multiple layers of features from tiny images,

    A. Krizhevsky and G. Hinton, “Learning multiple layers of features from tiny images,” University of Toronto, Toronto, Ontario, Tech. Rep., 2009

    2009

  10. [18]

    M2m: Imbalanced classification via major-to-minor translation,

    J. Kim, J. Jeong, and J. Shin, “M2m: Imbalanced classification via major-to-minor translation,” inProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 13 896–13 905

    2020

  11. [19]

    Large- scale long-tailed recognition in an open world,

    Z. Liu, Z. Miao, X. Zhan, J. Wang, B. Gong, and S. X. Yu, “Large- scale long-tailed recognition in an open world,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 2537–2546

    2019

  12. [20]

    ℓ p-norm support vector data description,

    S. Rahimzadeh Arashloo, “ℓ p-norm support vector data description,” Pattern Recognition, vol. 132, p. 108930, 2022

    2022

  13. [21]

    One-class classification usingℓ p-norm mul- tiple kernel fisher null approach,

    S. Rahimzadeh Arashloo, “One-class classification usingℓ p-norm mul- tiple kernel fisher null approach,”IEEE Transactions on Image Process- ing, vol. 32, pp. 1843–1856, 2023

    2023

  14. [22]

    Sta- bilizing adversarially learned one-class novelty detection using pseudo anomalies,

    M. Z. Zaheer, J.-H. Lee, A. Mahmood, M. Astrid, and S.-I. Lee, “Sta- bilizing adversarially learned one-class novelty detection using pseudo anomalies,”IEEE Transactions on Image Processing, vol. 31, pp. 5963– 5975, 2022

    2022

  15. [23]

    Target be- fore shooting: Accurate anomaly detection and localization under one millisecond via cascade patch retrieval,

    H. Li, J. Hu, B. Li, H. Chen, Y . Zheng, and C. Shen, “Target be- fore shooting: Accurate anomaly detection and localization under one millisecond via cascade patch retrieval,”IEEE Transactions on Image Processing, vol. 33, pp. 5606–5621, 2024

    2024

  16. [24]

    Drocc: Deep robust one-class classification,

    S. Goyal, A. Raghunathan, P. Jain, H. V . Simhadri, and P. N. Jain, “Drocc: Deep robust one-class classification,” inProceedings of the 37th International Conference on Machine Learning (ICML), ser. Proceedings of Machine Learning Research, vol. 119, 2020, pp. 3711–3721

    2020

  17. [25]

    Ocmst: One-class novelty detection using convolutional neural network and minimum spanning trees,

    R. La Grassa, I. Gallo, and N. Landro, “Ocmst: One-class novelty detection using convolutional neural network and minimum spanning trees,”Pattern Recognition Letters, vol. 155, pp. 114–120, 2022

    2022

  18. [26]

    Inter- pretable maximum margin deep anomaly detection,

    Z. Yang, M. Huang, X. Li, X. Pan, Q. Wang, and J. Zhao, “Inter- pretable maximum margin deep anomaly detection,”arXiv preprint arXiv:2603.07073, 2026

    2026

  19. [27]

    Csi: Novelty detection via contrastive learning on distributionally shifted instances,

    J. Tack, S. Mo, J. Jeong, and J. Shin, “Csi: Novelty detection via contrastive learning on distributionally shifted instances,” inAdvances in Neural Information Processing Systems, vol. 33, 2020

    2020

  20. [28]

    Deep anomaly de- tection with outlier exposure,

    D. Hendrycks, M. Mazeika, and T. G. Dietterich, “Deep anomaly de- tection with outlier exposure,” inInternational Conference on Learning Representations (ICLR), 2019

    2019

  21. [29]

    Universal novelty detection through adaptive contrastive learning,

    H. Mirzaei, M. Nafez, M. Jafari, M. B. Soltani, M. Azizmalayeri, J. Habibi, M. Sabokrou, and M. H. Rohban, “Universal novelty detection through adaptive contrastive learning,” inProceedings of the IEEE/CVF 13 Conference on Computer Vision and Pattern Recognition (CVPR), 2024, ...

    2024

  22. [30]

    Learning in-distribution representations for anomaly detection,

    W. T. Lunardi, A. Banabila, D. Herzalla, and M. L. Andreoni, “Learning in-distribution representations for anomaly detection,” 2025

    2025

  23. [31]

    Anomalyclip: Object- agnostic prompt learning for zero-shot anomaly detection,

    Q. Zhou, G. Pang, Y . Tian, S. He, and J. Chen, “Anomalyclip: Object- agnostic prompt learning for zero-shot anomaly detection,” inInterna- tional Conference on Learning Representations (ICLR), 2024

    2024

  24. [32]

    Aa-clip: Enhancing zero-shot anomaly detection via anomaly-aware clip,

    W. Ma, X. Zhang, Q. Yao, F. Tang, C. Wu, Y . Li, R. Yan, Z. Jiang, and S. K. Zhou, “Aa-clip: Enhancing zero-shot anomaly detection via anomaly-aware clip,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 4744– 4754

    2025

  25. [33]

    Fever-ood: Free energy vulnerability elimination for robust out-of-distribution detection,

    B. K. S. Isaac-Medina, M. Che, Y . F. A. Gaus, S. Akcay, and T. P. Breckon, “Fever-ood: Free energy vulnerability elimination for robust out-of-distribution detection,” inProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision (ICCV), 2025, pp. 4529–4538

    2025

  26. [34]

    Dual energy-based model with open-world un- certainty estimation for out-of-distribution detection,

    Q. Chen and H. Ding, “Dual energy-based model with open-world un- certainty estimation for out-of-distribution detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 25 728–25 737

    2025

  27. [35]

    Oodd: Test-time out-of-distribution detection with dynamic dictionary,

    Y . Yang, L. Zhu, Z. Sun, H. Liu, Q. Gu, and N. Ye, “Oodd: Test-time out-of-distribution detection with dynamic dictionary,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 30 630–30 639

    2025

  28. [36]

    Model-free test time adaptation for out-of-distribution detection,

    Y .-F. Zhang, X. Wang, T. Zhou, K. Yuan, Z. Zhang, L. Wang, and R. Jin, “Model-free test time adaptation for out-of-distribution detection,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 48, no. 2, pp. 1542–1553, 2026

    2026

  29. [37]

    Decoupling representation and classifier for long-tailed recognition,

    B. Kang, S. Xie, M. Rohrbach, Z. Yan, A. Gordo, J. Feng, and Y . Kalantidis, “Decoupling representation and classifier for long-tailed recognition,” inInternational Conference on Learning Representations (ICLR), 2020

    2020

  30. [38]

    Improving calibration for long-tailed recognition,

    Z. Zhong, J. Cui, S. Liu, and J. Jia, “Improving calibration for long-tailed recognition,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021, pp. 16 489–16 498

    2021

  31. [39]

    Reslt: Residual learning for long-tailed recognition,

    J. Cui, S. Liu, Z. Tian, Z. Zhong, and J. Jia, “Reslt: Residual learning for long-tailed recognition,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 3, pp. 3695–3706, 2023

    2023

  32. [40]

    Supervised exploratory learning for long-tailed visual recognition,

    Z. Jian, Y . Chen, Y . Wang, J. Yao, M. Wang, and Q. Wu, “Supervised exploratory learning for long-tailed visual recognition,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2025, pp. 1870–1880

    2025

  33. [41]

    Bce3s: Binary cross-entropy based tripartite synergistic learning for long-tailed recognition,

    W. Fan, Q. Li, J. Wen, and X. Peng, “Bce3s: Binary cross-entropy based tripartite synergistic learning for long-tailed recognition,” inProceedings of the AAAI Conference on Artificial Intelligence, 2026

    2026

  34. [42]

    Out-of-distribution de- tection in long-tailed recognition with calibrated outlier class learning,

    W. Miao, G. Pang, X. Bai, T. Li, and J. Zheng, “Out-of-distribution de- tection in long-tailed recognition with calibrated outlier class learning,” inProceedings of the AAAI Conference on Artificial Intelligence (AAAI), vol. 38, no. 5, 2024, pp. 4216–4224

    2024

  35. [43]

    Eat: Towards long-tailed out- of-distribution detection,

    T. Wei, B.-L. Wang, and M.-L. Zhang, “Eat: Towards long-tailed out- of-distribution detection,” inProceedings of the AAAI Conference on Artificial Intelligence (AAAI), vol. 38, no. 14, 2024, pp. 15 787–15 795

    2024

  36. [44]

    Rethinking out-of-distribution detection on imbalanced data distribution,

    K. Liu, Z. Fu, S. Jin, C. Chen, Z. Chen, R. Jiang, F. Zhou, Y . Chen, and J. Ye, “Rethinking out-of-distribution detection on imbalanced data distribution,” inAdvances in Neural Information Processing Systems, vol. 37, 2024

    2024

  37. [45]

    Long-tailed out-of-distribution detection via normalized outlier distribution adaptation,

    W. Miao, G. Pang, J. Zheng, and X. Bai, “Long-tailed out-of-distribution detection via normalized outlier distribution adaptation,” inAdvances in Neural Information Processing Systems, vol. 37, 2024

    2024

  38. [46]

    Revisiting Frank-Wolfe: Projection-free sparse convex op- timization,

    M. Jaggi, “Revisiting Frank-Wolfe: Projection-free sparse convex op- timization,” inProceedings of the 30th International Conference on Machine Learning, 2013, pp. 427–435

    2013

  39. [47]

    Nesterov,Lectures on Convex Optimization, 2nd ed

    Y . Nesterov,Lectures on Convex Optimization, 2nd ed. Springer, 2018

    2018

  40. [48]

    Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms,

    H. Xiao, K. Rasul, and R. V ollgraf, “Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms,”arXiv preprint arXiv:1708.07747, 2017

    2017 arXiv

  41. [49]

    On the use of a friedman-type statistic in balanced and unbalanced block designs,

    J. H. Skillings and G. A. Mack, “On the use of a friedman-type statistic in balanced and unbalanced block designs,”Technometrics, vol. 23, no. 2, pp. 171–177, 1981

    1981

  42. [50]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 770–778

    2016

  43. [51]

    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,” 2021

    2021

  44. [52]

    Efficient anomaly detection using self-supervised multi-cue tasks,

    L. J ´ez´equel, N.-S. Vu, J. Beaudet, and A. Histace, “Efficient anomaly detection using self-supervised multi-cue tasks,”IEEE Transactions on Image Processing, vol. 32, pp. 807–821, 2023

    2023

  45. [53]

    Learning and evalu- ating representations for deep one-class classification,

    K. Sohn, C.-L. Li, J. Yoon, M. Jin, and T. Pfister, “Learning and evalu- ating representations for deep one-class classification,” inInternational Conference on Learning Representations (ICLR), 2021

    2021

  46. [54]

    Admm-srnet: Alternating direction method of multipliers based sparse representation network for one-class classification,

    C.-Y . Chiou, K.-T. Lee, C.-R. Huang, and P.-C. Chung, “Admm-srnet: Alternating direction method of multipliers based sparse representation network for one-class classification,”IEEE Transactions on Image Processing, vol. 32, pp. 3447–3460, 2023

    2023

  47. [55]

    Latent outlier exposure for anomaly detection with contaminated data,

    C. Qiu, A. Li, M. Kloft, M. Rudolph, and S. Mandt, “Latent outlier exposure for anomaly detection with contaminated data,” inProceedings of the 39th International Conference on Machine Learning (ICML), ser. Proceedings of Machine Learning Research, vol. 162, 2022, pp. 18 153– 18 167

    2022

  48. [56]

    Uni- laterally aggregated contrastive learning with hierarchical augmentation for anomaly detection,

    G. Wang, Y . Wang, J. Qin, D. Zhang, X. Bao, and D. Huang, “Uni- laterally aggregated contrastive learning with hierarchical augmentation for anomaly detection,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023

    2023

  49. [57]

    Rodeo: Robust outlier detection via exposing adaptive out-of- distribution samples,

    H. Mirzaei, M. Jafari, H. R. Dehbashi, A. Ansari, S. Ghobadi, M. Hadi, A. S. Moakhar, M. Azizmalayeri, M. S. Baghshah, and M. H. Ro- hban, “Rodeo: Robust outlier detection via exposing adaptive out-of- distribution samples,”arXiv preprint arXiv:2501.16971, 2025

    2025 arXiv

  50. [58]

    An evidence-based post-hoc adjustment framework for anomaly detection under data contamination,

    S. Patra and S. Ben Taieb, “An evidence-based post-hoc adjustment framework for anomaly detection under data contamination,”arXiv preprint arXiv:2510.21296, 2025

    2025

  51. [59]

    Self supervision to distillation for long- tailed visual recognition,

    T. Li, L. Wang, and G. Wu, “Self supervision to distillation for long- tailed visual recognition,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021, pp. 630–639

    2021

  52. [60]

    Weight balancing for long- tailed image classification,

    B. Zhao, C. Li, Y . Li, J. Wang, and Y . Fu, “Weight balancing for long- tailed image classification,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 6897– 6907

    2023

  53. [61]

    Learning from multiple experts: Self- paced knowledge distillation for long-tailed classification,

    L. Xiang, G. Ding, and J. Han, “Learning from multiple experts: Self- paced knowledge distillation for long-tailed classification,” inProceed- ings of the European Conference on Computer Vision (ECCV), 2020, pp. 247–263

    2020

  54. [62]

    Parametric contrastive learning,

    J. Cui, Z. Zhong, S. Liu, B. Yu, and J. Jia, “Parametric contrastive learning,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021, pp. 715–724

    2021

  55. [63]

    Balanced contrastive learning for long-tailed visual recognition,

    J.-X. Zhu, T. Wang, Z. Chen, J.-J. Liu, M.-M. Ye, and X.-S. Cheng, “Balanced contrastive learning for long-tailed visual recognition,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 6908–6917

    2022

  56. [64]

    Diffult: Diffusion for long- tail recognition without external knowledge,

    J. Shao, K. Zhu, H. Zhang, and J. Wu, “Diffult: Diffusion for long- tail recognition without external knowledge,” inAdvances in Neural Information Processing Systems, 2024

    2024

  57. [65]

    Ltrl: Boosting long-tail recognition via reflective learning,

    Q. Zhao, Y . Dai, S. Lin, W. Hu, F. Zhang, and J. Liu, “Ltrl: Boosting long-tail recognition via reflective learning,” inProceedings of the European Conference on Computer Vision, 2024

    2024

  58. [66]

    Focal-sam: Focal sharpness-aware minimization for long-tailed classi- fication,

    S. Li, Q. Xu, Z. Yang, Z. Wang, L. Zhang, X. Cao, and Q. Huang, “Focal-sam: Focal sharpness-aware minimization for long-tailed classi- fication,” inProceedings of the International Conference on Machine Learning, 2025

    2025

  59. [67]

    Long-tailed classification with multi-granularity seman- tics,

    Y . Liuet al., “Long-tailed classification with multi-granularity seman- tics,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025

    2025

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.