pith. sign in

arxiv: 2606.04920 · v1 · pith:L445IBJCnew · submitted 2026-06-03 · 💻 cs.LG · cs.CV

Toward Multi-Domain and Long-Tailed Quantization via Feature Alignment and Scaling

Chin-Yuan Yeh , Ting-An Chen , De-Nian Yang , Ming-Syan Chen This is my paper

Pith reviewed 2026-06-28 07:28 UTC · model grok-4.3

classification 💻 cs.LG cs.CV
keywords quantizationmulti-domain learninglong-tailed distributionfeature alignmentclass imbalancedeep neural networkslow-bit inferencedomain shift
0
0 comments X

The pith

EmaQ aligns multi-domain distributions with CDF projection and sensitivity-aware aggregation to enable effective low-bit quantization under shifts and imbalance.

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

The paper introduces Efficient Multi-Domain Alignment Quantization (EmaQ) to quantize deep neural networks when training data spans multiple domains or exhibits severe class imbalance. It projects features onto a common distribution via their cumulative distribution functions and aggregates weights according to sensitivity to produce stable quantized models. An extension called EmaQ-LT adds class-conditioned variance scaling and confidence-based logit adjustment to counter majority-class dominance. Theoretical analysis supplies convergence guarantees for these mechanisms. Experiments across standard, multi-domain, and long-tailed benchmarks demonstrate improved accuracy at low bit widths compared with prior quantization techniques.

Core claim

The authors establish that CDF-based feature projection aligns domain distributions while sensitivity-aware weight aggregation stabilizes the quantization process, and that extending these with class-conditioned variance scaling and logit adjustment produces reliable low-bit models for both multi-domain and long-tailed settings, backed by convergence proofs and benchmark results on Office-31, Digits, SynDigits-LT, CIFAR-10-LT, and CIFAR-100-LT.

What carries the argument

CDF-based projection for domain alignment together with sensitivity-aware weight aggregation, extended by class-conditioned variance scaling and confidence-based logit adjustment.

If this is right

  • Quantized models can be trained once on mixed-domain data and deployed across domains without retraining.
  • Low-bit networks maintain accuracy on long-tailed datasets by reducing majority-class overconfidence.
  • Convergence of the quantization process is guaranteed when the sensitivity and scaling terms are used.
  • The same alignment and scaling steps apply directly to other low-bit widths such as 2-bit or 3-bit.

Where Pith is reading between the lines

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

  • The same CDF alignment step might transfer to other compression methods such as pruning or knowledge distillation when domain shift is present.
  • If the projection truly preserves task-relevant features, quantized models could require less domain-specific adaptation at deployment time.
  • Combining EmaQ with online domain detection would allow the method to handle streaming data whose domain distribution changes over time.

Load-bearing premise

The CDF projection aligns domains while keeping information needed for accurate quantization, and the sensitivity aggregation plus class scaling introduce no unaccounted bias and do not require domain labels at test time.

What would settle it

Running EmaQ on Office-31 or CIFAR-10-LT at 4-bit width and finding no accuracy gain over standard post-training quantization baselines would falsify the central performance claim.

Figures

Figures reproduced from arXiv: 2606.04920 by Chin-Yuan Yeh, De-Nian Yang, Ming-Syan Chen, Ting-An Chen.

Figure 1
Figure 1. Figure 1: Motivation for feature scaling. (a) Large quantization errors arise when source and target domains diverge. (b) Uneven source-domain sensitivities cause inconsistent con￾vergence; sensitive domains can be distorted by uniform aggregation, degrading the aggregated result (red) and overall quantization performance [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Motivation for logit scaling. Prediction probabilities under 4-bit quantization on a long-tailed dataset with C = 10 classes. A majority class (e.g., class 1) attains substantially higher confidence than minority classes (e.g., class 9), moti￾vating the need for logit adjustment. cross-domain variation or imbalanced class frequencies. For instance, images captured on different devices can lead to domain sh… view at source ↗
Figure 3
Figure 3. Figure 3: Motivation for variance and logit scaling. (a) A long-tailed distribution of C = 10 classes. (b) Long-tailed quantization causes diverse drops in accuracy. (c) T-SNE plot of class features shows that classes exhibit significantly different variances due to unequal sample sizes. EmaQ includes Domain Alignment Quantization Training (DAQT), which projects features from each domain onto a common uniform distri… view at source ↗
Figure 4
Figure 4. Figure 4: CNN weight and activation distributions for ResNet [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Layer-wise quantization errors without/with class [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Convergence of AQGD compared against SGD. Three figures present the losses on (a) Amazon, (b) DSLR and (c) [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Visualization of class weights in the homogenized loss [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
read the original abstract

Quantizing deep neural networks is essential for efficient inference on resource-constrained devices. However, most existing methods are designed for single-domain and class-balanced data, leaving practical settings with domain shifts or severe class imbalance underexplored. We address these challenges with Efficient Multi-Domain Alignment Quantization (EmaQ), which aligns domain distributions through a CDF-based projection and uses sensitivity-aware weight aggregation to stabilize multi-domain quantization. We further extend EmaQ to EmaQ-LT for long-tailed quantization by introducing class-conditioned variance scaling and confidence-based logit adjustment to mitigate majority-class overconfidence. Theoretical analyses establish convergence guarantees and motivate the proposed sensitivity and scaling mechanisms. Experiments on standard, multi-domain (Office-31, Digits), and long-tailed (SynDigits-LT, CIFAR-10-LT, CIFAR-100-LT) benchmarks show that EmaQ and EmaQ-LT achieve strong low-bit performance under domain shift and class imbalance.

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

3 major / 2 minor

Summary. The paper proposes Efficient Multi-Domain Alignment Quantization (EmaQ) that aligns domain distributions via a CDF-based projection and applies sensitivity-aware weight aggregation for stable multi-domain quantization. It extends the method to EmaQ-LT for long-tailed settings by adding class-conditioned variance scaling and confidence-based logit adjustment. Theoretical analyses are claimed to establish convergence guarantees, and experiments on Office-31, Digits, SynDigits-LT, CIFAR-10-LT, and CIFAR-100-LT are reported to show strong low-bit performance under domain shift and class imbalance.

Significance. If the central mechanisms are shown to preserve quantization-relevant statistics under the proposed transforms, the work would address an underexplored practical setting for quantization. The combination of alignment, sensitivity aggregation, and class-conditioned scaling is a coherent attempt to handle distribution mismatch and imbalance simultaneously.

major comments (3)
  1. [§3.2] §3.2 (CDF-based projection): the manuscript must demonstrate that the per-feature monotonic CDF transform leaves the relative ordering and magnitudes of Hessian-based sensitivities (or activation ranges) unchanged across domains; otherwise the subsequent sensitivity-aware aggregation step operates on misaligned statistics and the convergence analysis does not apply to the composed operator.
  2. [§4] §4 (theoretical analysis): the convergence guarantee is stated to motivate the sensitivity and scaling mechanisms, yet it is unclear whether the proof accounts for the bias potentially introduced by the CDF projection or the class-conditioned scaling; an explicit statement of the assumptions under which the guarantee holds after these transforms is required.
  3. [Table 2] Table 2 (multi-domain results): the reported gains for EmaQ over single-domain baselines lack error bars and details on the number of random seeds; without this, it is impossible to assess whether the improvement under domain shift is statistically reliable or sensitive to post-hoc hyper-parameter choices.
minor comments (2)
  1. [§3] Notation for the sensitivity metric (e.g., Eq. (X)) should be defined before its first use in the method section rather than introduced inline.
  2. [§5.2] The long-tailed benchmarks (SynDigits-LT, CIFAR-*-LT) should include a reference or explicit description of how the imbalance ratios were constructed to allow reproduction.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major comment point-by-point below, indicating where revisions will be made to strengthen the paper.

read point-by-point responses

  1. Referee: [§3.2] §3.2 (CDF-based projection): the manuscript must demonstrate that the per-feature monotonic CDF transform leaves the relative ordering and magnitudes of Hessian-based sensitivities (or activation ranges) unchanged across domains; otherwise the subsequent sensitivity-aware aggregation step operates on misaligned statistics and the convergence analysis does not apply to the composed operator.

    Authors: We thank the referee for this observation. The per-feature CDF transform is strictly monotonic, which preserves the relative ordering of activation values within each feature. To address the concern regarding Hessian-based sensitivities, we will add a supporting argument and lemma in the revised §3.2 (and appendix) showing that, because the transform is applied uniformly after domain alignment, the relative magnitudes of per-feature sensitivities remain consistent across domains. This ensures the sensitivity-aware aggregation operates on aligned statistics. revision: yes

  2. Referee: [§4] §4 (theoretical analysis): the convergence guarantee is stated to motivate the sensitivity and scaling mechanisms, yet it is unclear whether the proof accounts for the bias potentially introduced by the CDF projection or the class-conditioned scaling; an explicit statement of the assumptions under which the guarantee holds after these transforms is required.

    Authors: We agree that the assumptions require explicit clarification. In the revised §4, we will update the theorem statement and proof sketch to list the assumptions under which the guarantee holds after the transforms: specifically, that the CDF projection is a strictly increasing continuous bijection and that the class-conditioned scaling factors are positive and bounded. Under these conditions, the bias is controlled and the contraction property used in the convergence analysis is preserved for the composed operator. revision: yes

  3. Referee: [Table 2] Table 2 (multi-domain results): the reported gains for EmaQ over single-domain baselines lack error bars and details on the number of random seeds; without this, it is impossible to assess whether the improvement under domain shift is statistically reliable or sensitive to post-hoc hyper-parameter choices.

    Authors: We acknowledge the importance of reporting statistical reliability. In the revised manuscript, we will update Table 2 to include error bars (standard deviations) computed over multiple random seeds and explicitly state the number of seeds used for the multi-domain experiments. This will allow readers to better evaluate the robustness of the reported gains. revision: yes

Circularity Check

0 steps flagged

No circularity: derivation self-contained against external benchmarks

full rationale

The provided abstract and excerpts describe CDF projection, sensitivity-aware aggregation, class-conditioned scaling, and convergence analyses without any quoted equations or self-citations that reduce a claimed prediction or uniqueness result to a fitted input or prior author work by construction. Theoretical guarantees are presented as independent motivation rather than tautological. No load-bearing self-citation chain or renaming of known results is exhibited in the given text. This matches the common honest finding of a self-contained paper.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities; no equations or detailed derivations available to audit.

pith-pipeline@v0.9.1-grok · 5707 in / 985 out tokens · 27392 ms · 2026-06-28T07:28:49.988660+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

98 extracted references · 6 canonical work pages · 2 internal anchors

  1. [1]

    A survey of quantization methods for efficient neural network infer- ence,

    A. Gholami, S. Kim, Z. Dong, Z. Yao, M. W. Mahoney, and K. Keutzer, “A survey of quantization methods for efficient neural network infer- ence,” inLow-Power Computer Vision. Chapman and Hall/CRC, 2022, pp. 291–326

    2022

  2. [2]

    Quantization networks,

    J. Yang, X. Shen, J. Xing, X. Tian, H. Li, B. Deng, J. Huang, and X.-s. Hua, “Quantization networks,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 7308– 7316

    2019

  3. [3]

    Quantization of deep neural networks for accurate edge computing,

    W. Chen, H. Qiu, J. Zhuang, C. Zhang, Y . Hu, Q. Lu, T. Wang, Y . Shi, M. Huang, and X. Xu, “Quantization of deep neural networks for accurate edge computing,”ACM Journal on Emerging Technologies in Computing Systems (JETC), vol. 17, no. 4, pp. 1–11, 2021

    2021

  4. [4]

    Deep learning with low precision by half-wave gaussian quantization,

    Z. Cai, X. He, J. Sun, and N. Vasconcelos, “Deep learning with low precision by half-wave gaussian quantization,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 5918–5926

    2017

  5. [5]

    Learned step size quantization,

    S. K. Esser, J. L. McKinstry, D. Bablani, R. Appuswamy, and D. S. Modha, “Learned step size quantization,” inInternational Conference on Learning Representations, 2020. [Online]. Available: https://openreview.net/forum?id=rkgO66VKDS

    2020

  6. [6]

    Additive powers-of-two quantization: An efficient non-uniform discretization for neural networks,

    Y . Li, X. Dong, and W. Wang, “Additive powers-of-two quantization: An efficient non-uniform discretization for neural networks,” in International Conference on Learning Representations, 2020. [Online]. Available: https://openreview.net/forum?id=BkgXT24tDS

    2020

  7. [8]

    Data-free network quanti- zation with adversarial knowledge distillation,

    Y . Choi, J. Choi, M. El-Khamy, and J. Lee, “Data-free network quanti- zation with adversarial knowledge distillation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2020, pp. 710–711

    2020

  8. [9]

    Generative zero- shot network quantization,

    X. He, J. Lu, W. Xu, Q. Hu, P. Wang, and J. Cheng, “Generative zero- shot network quantization,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 3000–3011

    2021

  9. [11]

    A study on real-time image processing applications with edge computing support for mobile devices,

    G. P. Mattia and R. Beraldi, “A study on real-time image processing applications with edge computing support for mobile devices,” in2021 IEEE/ACM 25th International Symposium on Distributed Simulation and Real Time Applications (DS-RT). IEEE, 2021, pp. 1–7

    2021

  10. [12]

    Color correction of the document owner’s photograph image during recognition on mobile device,

    D. Polevoy, E. Panfilova, E. Ershov, and D. Nikolaev, “Color correction of the document owner’s photograph image during recognition on mobile device,” inThirteenth International Conference on Machine Vision, vol. 11605. SPIE, 2021, pp. 285–293

    2021

  11. [13]

    Lote-animal: A long time-span dataset for endangered animal behavior understanding,

    D. Liu, J. Hou, S. Huang, J. Liu, Y . He, B. Zheng, J. Ning, and J. Zhang, “Lote-animal: A long time-span dataset for endangered animal behavior understanding,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 20 064–20 075

    2023

  12. [14]

    Zeroq: A novel zero shot quantization framework,

    Y . Cai, Z. Yao, Z. Dong, A. Gholami, M. W. Mahoney, and K. Keutzer, “Zeroq: A novel zero shot quantization framework,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 13 169–13 178

    2020

  13. [15]

    Zero-shot adversarial quantization,

    Y . Liu, W. Zhang, and J. Wang, “Zero-shot adversarial quantization,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 1512–1521

    2021

  14. [16]

    Aciq: Analytical clipping for integer quantization of neural networks,

    R. Banner, Y . Nahshan, E. Hoffer, and D. Soudry, “Aciq: Analytical clipping for integer quantization of neural networks,” 2018

    2018

  15. [17]

    Data-free quantization via pseudo-label filtering,

    C. Fan, Z. Wang, D. Guo, and M. Wang, “Data-free quantization via pseudo-label filtering,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 5589–5598

    2024

  16. [18]

    Metaaug: meta-data augmentation for post-training quan- tization,

    C. Pham, A. D. Hoang, C. C. Nguyen, T. Le, D. Phung, G. Carneiro, and T.-T. Do, “Metaaug: meta-data augmentation for post-training quan- tization,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 236–252

    2024

  17. [19]

    Genie: show me the data for quantization,

    Y . Jeon, C. Lee, and H.-y. Kim, “Genie: show me the data for quantization,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 12 064–12 073

    2023

  18. [20]

    Pd-quant: Post-training quantization based on prediction difference metric,

    J. Liu, L. Niu, Z. Yuan, D. Yang, X. Wang, and W. Liu, “Pd-quant: Post-training quantization based on prediction difference metric,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 24 427–24 437

    2023

  19. [21]

    I&s-vit: An inclusive & stable method for post-training vits quantization,

    Y . Zhong, J. Hu, M. Lin, M. Chen, and R. Ji, “I&s-vit: An inclusive & stable method for post-training vits quantization,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025

  20. [22]

    Instance-aware group quantization for vision transformers,

    J. Moon, D. Kim, J. Cheon, and B. Ham, “Instance-aware group quantization for vision transformers,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 16 132–16 141

    2024

  21. [23]

    Towards accu- rate post-training quantization for diffusion models,

    C. Wang, Z. Wang, X. Xu, Y . Tang, J. Zhou, and J. Lu, “Towards accu- rate post-training quantization for diffusion models,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 16 026–16 035

    2024

  22. [24]

    Aphq- vit: Post-training quantization with average perturbation hessian based reconstruction for vision transformers,

    Z. Wu, J. Zhang, J. Chen, J. Guo, D. Huang, and Y . Wang, “Aphq- vit: Post-training quantization with average perturbation hessian based reconstruction for vision transformers,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 9686–9695

    2025

  23. [25]

    Enhancing post-training quantization calibration through contrastive learning,

    Y . Shang, G. Liu, R. R. Kompella, and Y . Yan, “Enhancing post-training quantization calibration through contrastive learning,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 15 921–15 930

    2024

  24. [26]

    Magr: Weight magnitude reduction for enhancing post-training quantization,

    A. Zhang, N. Wang, Y . Deng, X. Li, Z. Yang, and P. Yin, “Magr: Weight magnitude reduction for enhancing post-training quantization,”Advances in neural information processing systems, vol. 37, pp. 85 109–85 130, 2024

    2024

  25. [27]

    Towards next-level post-training quantization of hyper-scale transform- ers,

    J. Kim, C. Lee, E. Cho, K. Park, H.-y. Kim, J. Kim, and Y . Jeon, “Towards next-level post-training quantization of hyper-scale transform- ers,”Advances in Neural Information Processing Systems, vol. 37, pp. 94 292–94 326, 2024

    2024

  26. [28]

    Additive powers-of-two quantization: An efficient non-uniform discretization for neural networks,

    Y . Li, X. Dong, and W. Wang, “Additive powers-of-two quantization: An efficient non-uniform discretization for neural networks,” in IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 0, NO. 0, NOVEMBER 2022 13 International Conference on Learning Representations, 2020. [Online]. Available: https://openreview.net/forum?id=BkgXT24tDS

    2022

  27. [29]

    Quantization without tears,

    M. Fu, H. Yu, J. Shao, J. Zhou, K. Zhu, and J. Wu, “Quantization without tears,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 4462–4472

    2025

  28. [30]

    Efficient multi-bit quantization network training via weight bias correction and bit- wise coreset sampling,

    J. Kim, J. J. An, K. E. Jeon, and J. H. Ko, “Efficient multi-bit quantization network training via weight bias correction and bit- wise coreset sampling,” inThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2026. [Online]. Available: https://openreview.net/forum?id=VBKgukQlRG

    2026

  29. [31]

    Eq-net: Elastic quanti- zation neural networks,

    K. Xu, L. Han, Y . Tian, S. Yang, and X. Zhang, “Eq-net: Elastic quanti- zation neural networks,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 1505–1514

    2023

  30. [32]

    Automatic joint structured pruning and quantization for efficient neural network training and compression,

    X. Qu, D. Aponte, C. Banbury, D. P. Robinson, T. Ding, K. Koishida, I. Zharkov, and T. Chen, “Automatic joint structured pruning and quantization for efficient neural network training and compression,” in Proceedings of the Computer Vision and Pattern Recognition Confer- ence, 2025, pp. 15 234–15 244

    2025

  31. [33]

    Domain-adversarial training of neural networks,

    Y . Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Lavi- olette, M. March, and V . Lempitsky, “Domain-adversarial training of neural networks,”Journal of machine learning research, vol. 17, no. 59, pp. 1–35, 2016

    2016

  32. [34]

    Intraq: Learning synthetic images with intra-class heterogeneity for zero-shot network quantization,

    Y . Zhong, M. Lin, G. Nan, J. Liu, B. Zhang, Y . Tian, and R. Ji, “Intraq: Learning synthetic images with intra-class heterogeneity for zero-shot network quantization,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 12 339–12 348

    2022

  33. [35]

    Towards better robust- ness against common corruptions for unsupervised domain adaptation,

    Z. Gao, K. Huang, R. Zhang, D. Liu, and J. Ma, “Towards better robust- ness against common corruptions for unsupervised domain adaptation,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 18 882–18 893

    2023

  34. [36]

    Qt-dog: Quantization-aware training for domain generalization,

    S. Javed, H. Le, and M. Salzmann, “Qt-dog: Quantization-aware training for domain generalization,”arXiv preprint arXiv:2410.06020, 2024

    work page arXiv 2024

  35. [37]

    Gaqat: gradient-adaptive quantization-aware training for domain gen- eralization,

    J. Jiang, Y . Meng, C. Tang, H. Yu, Q. Li, Z. Wang, and W. Zhu, “Gaqat: gradient-adaptive quantization-aware training for domain gen- eralization,”arXiv preprint arXiv:2412.05551, 2024

    work page arXiv 2024

  36. [38]

    Alignq: Alignment quan- tization with admm-based correlation preservation,

    T.-A. Chen, D.-N. Yang, and M.-S. Chen, “Alignq: Alignment quan- tization with admm-based correlation preservation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 12 538–12 547

    2022

  37. [39]

    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, 2019, pp. 2537–2546

    2019

  38. [40]

    Deep long-tailed learning: A survey,

    Y . Zhang, B. Kang, B. Hooi, S. Yan, and J. Feng, “Deep long-tailed learning: A survey,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023

    2023

  39. [41]

    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,”arXiv preprint arXiv:1910.09217, 2019

    work page arXiv 1910

  40. [42]

    Dynamic curriculum learning for imbalanced data classification,

    Y . Wang, W. Gan, J. Yang, W. Wu, and J. Yan, “Dynamic curriculum learning for imbalanced data classification,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 5017– 5026

    2019

  41. [43]

    Rethinking classifier re-training in long-tailed recognition: Label over-smooth can balance,

    S. Sun, H. Lu, J. Li, Y . Xie, T. Li, X. Yang, L. Zhang, and J. Yan, “Rethinking classifier re-training in long-tailed recognition: Label over-smooth can balance,” inThe Thirteenth International Conference on Learning Representations, 2025. [Online]. Available: https://openreview.net/forum?id=OeKp3AdiVO

    2025

  42. [44]

    Ltgc: Long- tail recognition via leveraging llms-driven generated content,

    Q. Zhao, Y . Dai, H. Li, W. Hu, F. Zhang, and J. Liu, “Ltgc: Long- tail recognition via leveraging llms-driven generated content,” inPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 19 510–19 520

    2024

  43. [45]

    Breaking long-tailed learning bottlenecks: A controllable paradigm with hypernetwork-generated diverse experts,

    Z. Zhao, H. Wen, Z. Wang, P. Wang, F. Wang, S. Lai, Q. Zhang, and Y . Wang, “Breaking long-tailed learning bottlenecks: A controllable paradigm with hypernetwork-generated diverse experts,”Advances in Neural Information Processing Systems, vol. 37, pp. 7493–7520, 2024

    2024

  44. [46]

    Disentangling label distribution for long-tailed visual recognition,

    Y . Hong, S. Han, K. Choi, S. Seo, B. Kim, and B. Chang, “Disentangling label distribution for long-tailed visual recognition,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 6626–6636

    2021

  45. [47]

    Overcoming forgetting catas- trophe in quantization-aware training,

    T.-A. Chen, D.-N. Yang, and M.-S. Chen, “Overcoming forgetting catas- trophe in quantization-aware training,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 17 358–17 367

    2023

  46. [48]

    Lightlt: a lightweight representation quantization framework for long-tail data,

    H. Wang, R. Li, Z. Wang, X. Tang, D. Zhang, M. Cheng, B. Yin, J. Droppo, S. Wang, and J. Gao, “Lightlt: a lightweight representation quantization framework for long-tail data,” in2024 IEEE 40th Inter- national Conference on Data Engineering (ICDE). IEEE, 2024, pp. 1380–1393

    2024

  47. [49]

    Metaquant: Learning to quantize by learning to penetrate non-differentiable quantization,

    S. Chen, W. Wang, and S. J. Pan, “Metaquant: Learning to quantize by learning to penetrate non-differentiable quantization,”Advances in Neural Information Processing Systems, vol. 32, 2019

    2019

  48. [50]

    Adaptive proximal gradient meth- ods for structured neural networks,

    J. Yun, A. C. Lozano, and E. Yang, “Adaptive proximal gradient meth- ods for structured neural networks,”Advances in Neural Information Processing Systems, vol. 34, pp. 24 365–24 378, 2021

    2021

  49. [51]

    Shalev-Shwartz and S

    S. Shalev-Shwartz and S. Ben-David,Understanding machine learning: From theory to algorithms. Cambridge university press, 2014

    2014

  50. [52]

    Levene tests of homogeneity of variance for general block and treatment designs,

    M. E. O’Neill and K. L. Mathews, “Levene tests of homogeneity of variance for general block and treatment designs,”Biometrics, vol. 58, no. 1, pp. 216–224, 2002

    2002

  51. [54]

    A tutorial on the cross-entropy method,

    P.-T. De Boer, D. P. Kroese, S. Mannor, and R. Y . Rubinstein, “A tutorial on the cross-entropy method,”Annals of operations research, vol. 134, pp. 19–67, 2005

    2005

  52. [55]

    Linear symmetric quantization of neural networks for low-precision integer hardware,

    X. Zhao, Y . Wang, X. Cai, C. Liu, and L. Zhang, “Linear symmetric quantization of neural networks for low-precision integer hardware,” in International Conference on Learning Representations, 2020

    2020

  53. [56]

    DoReFa-Net: Training Low Bitwidth Convolutional Neural Networks with Low Bitwidth Gradients

    S. Zhou, Y . Wu, Z. Ni, X. Zhou, H. Wen, and Y . Zou, “Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients,”arXiv preprint arXiv:1606.06160, 2016

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  54. [57]

    Improving neural network quantization without retraining using outlier channel splitting,

    R. Zhao, Y . Hu, J. Dotzel, C. De Sa, and Z. Zhang, “Improving neural network quantization without retraining using outlier channel splitting,” inInternational Conference on Machine Learning. PMLR, 2019, pp. 7543–7552

    2019

  55. [58]

    Generative zero- shot network quantization,

    X. He, J. Lu, W. Xu, Q. Hu, P. Wang, and J. Cheng, “Generative zero- shot network quantization,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 3000–3011

    2021

  56. [59]

    Gener- ative low-bitwidth data free quantization,

    S. Xu, H. Li, B. Zhuang, J. Liu, J. Cao, C. Liang, and M. Tan, “Gener- ative low-bitwidth data free quantization,” inEuropean Conference on Computer Vision. Springer, 2020, pp. 1–17

    2020

  57. [60]

    It’s all in the teacher: Zero-shot quantization brought closer to the teacher,

    K. Choi, H. Y . Lee, D. Hong, J. Yu, N. Park, Y . Kim, and J. Lee, “It’s all in the teacher: Zero-shot quantization brought closer to the teacher,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 8311–8321

    2022

  58. [61]

    Rethinking data-free quan- tization as a zero-sum game,

    B. Qian, Y . Wang, R. Hong, and M. Wang, “Rethinking data-free quan- tization as a zero-sum game,” inProceedings of the AAAI conference on artificial intelligence, vol. 37, no. 8, 2023, pp. 9489–9497

    2023

  59. [62]

    Adaptive data-free quantization,

    ——, “Adaptive data-free quantization,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 7960–7968

    2023

  60. [63]

    Texq: Zero- shot network quantization with texture feature distribution calibration,

    X. Chen, Y . Wang, R. Yan, Y . Liu, T. Guan, and Y . He, “Texq: Zero- shot network quantization with texture feature distribution calibration,” Advances in Neural Information Processing Systems, vol. 36, pp. 274– 287, 2023

    2023

  61. [64]

    Qimera: Data-free quantization with synthetic boundary supporting samples,

    K. Choi, D. Hong, N. Park, Y . Kim, and J. Lee, “Qimera: Data-free quantization with synthetic boundary supporting samples,”Advances in Neural Information Processing Systems, vol. 34, 2021

    2021

  62. [65]

    Batchquant: Quantized- for-all architecture search with robust quantizer,

    H. Bai, M. Cao, P. Huang, and J. Shan, “Batchquant: Quantized- for-all architecture search with robust quantizer,”Advances in Neural Information Processing Systems, vol. 34, 2021

    2021

  63. [66]

    SQuant: On-the-fly data-free quantization via diagonal hessian approximation,

    C. Guo, Y . Qiu, J. Leng, X. Gao, C. Zhang, Y . Liu, F. Yang, Y . Zhu, and M. Guo, “SQuant: On-the-fly data-free quantization via diagonal hessian approximation,” inInternational Conference on Learning Representations, 2022. [Online]. Available: https://openreview.net/forum?id=JXhROKNZzOc

    2022

  64. [67]

    Learning multiple layers of features from tiny images,

    A. Krizhevsky, G. Hintonet al., “Learning multiple layers of features from tiny images,” 2009

    2009

  65. [68]

    Reading digits in natural images with unsupervised feature learning,

    Y . Netzer, T. Wang, A. Coates, A. Bissacco, B. Wu, and A. Y . Ng, “Reading digits in natural images with unsupervised feature learning,” 2011

    2011

  66. [69]

    Imagenet large scale visual recognition challenge,

    O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernsteinet al., “Imagenet large scale visual recognition challenge,”International journal of computer vision, vol. 115, no. 3, pp. 211–252, 2015

    2015

  67. [70]

    Adapting visual category models to new domains,

    K. Saenko, B. Kulis, M. Fritz, and T. Darrell, “Adapting visual category models to new domains,” inEuropean conference on computer vision. Springer, 2010, pp. 213–226

    2010

  68. [71]

    Gradient-based learning applied to document recognition,

    Y . LeCun, L. Bottou, Y . Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,”Proceedings of the IEEE, vol. 86, no. 11, pp. 2278–2324, 1998. IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 0, NO. 0, NOVEMBER 2022 14

    1998

  69. [72]

    Unsupervised domain adaptation by back- propagation,

    Y . Ganin and V . Lempitsky, “Unsupervised domain adaptation by back- propagation,” inInternational conference on machine learning. PMLR, 2015, pp. 1180–1189

    2015

  70. [73]

    Very Deep Convolutional Networks for Large-Scale Image Recognition

    K. Simonyan, “Very deep convolutional networks for large-scale image recognition,”arXiv preprint arXiv:1409.1556, 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  71. [74]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778

    2016

  72. [75]

    Densely connected convolutional networks,

    G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, “Densely connected convolutional networks,” inProceedings of the IEEE confer- ence on computer vision and pattern recognition, 2017, pp. 4700–4708

    2017

  73. [76]

    Mobilenetv2: Inverted residuals and linear bottlenecks,

    M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen, “Mobilenetv2: Inverted residuals and linear bottlenecks,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 4510–4520

    2018

  74. [77]

    Research in probability and statistics: Reflections and directions

    J. M. Shaughnessy, “Research in probability and statistics: Reflections and directions.” 1992

    1992

  75. [78]

    Up or down? adaptive rounding for post-training quantization,

    M. Nagel, R. A. Amjad, M. Van Baalen, C. Louizos, and T. Blankevoort, “Up or down? adaptive rounding for post-training quantization,” in International Conference on Machine Learning. PMLR, 2020, pp. 7197–7206

    2020

  76. [79]

    Rethinking the evaluation protocol of domain generalization,

    H. Yu, X. Zhang, R. Xu, J. Liu, Y . He, and P. Cui, “Rethinking the evaluation protocol of domain generalization,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 21 897–21 908

    2024

  77. [80]

    Learning frequency-adapted vision foundation model for domain generalized semantic segmentation,

    Q. Bi, J. Yi, H. Zheng, H. Zhan, Y . Huang, W. Ji, Y . Li, and Y . Zheng, “Learning frequency-adapted vision foundation model for domain generalized semantic segmentation,”Advances in Neural Infor- mation Processing Systems, vol. 37, pp. 94 047–94 072, 2024

    2024

  78. [81]

    Transferrable prototypical networks for unsupervised domain adaptation,

    Y . Pan, T. Yao, Y . Li, Y . Wang, C.-W. Ngo, and T. Mei, “Transferrable prototypical networks for unsupervised domain adaptation,” inPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 2239–2247

    2019

  79. [82]

    Contrastively smoothed class alignment for unsupervised domain adaptation,

    S. Dai, Y . Cheng, Y . Zhang, Z. Gan, J. Liu, and L. Carin, “Contrastively smoothed class alignment for unsupervised domain adaptation,” in Proceedings of the Asian conference on computer vision, 2020

    2020

  80. [83]

    Joint domain alignment and class alignment method for cross-domain fault diagnosis of rotating machinery,

    Y . Zhang, K. Yu, Z. Ren, and S. Zhou, “Joint domain alignment and class alignment method for cross-domain fault diagnosis of rotating machinery,”IEEE Transactions on Instrumentation and Measurement, vol. 70, pp. 1–12, 2021

    2021

Showing first 80 references.