Pith. sign in

REVIEW 4 major objections 6 minor 60 references

Quantization Meets OOD: Generalizable Quantization-aware Training from a Flatness Perspective

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read Standard quantization-aware training degrades a model's out-of-distribution accuracy by sharpening its loss landscape; a flatness objective with adaptive gradient freezing restores it.

desk verdict Useful method, overclaimed diagnosis: the OOD drop may be an artifact of missing long-training control. read the letter →

arxiv 2509.00859 v1 pith:3Y3KPAK7 submitted 2025-08-31 cs.CV

classification cs.CV
keywords quantization-awaretrainingout-of-distributiongeneralizationflatnessoflosslandscapegradientdisorderlayerfreezingsharpness-awareminimizationdomainscalefactor
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

Quantization-aware training (QAT) reduces the bit width of neural networks for on-device deployment, but this paper argues that QAT silently damages accuracy on data that differs from the training distribution — exactly the distribution shift edge deployments face. The authors measure two standard QAT methods on three domain-generalization benchmarks and report large OOD accuracy drops, while showing that an earlier conflicting result is likely an artifact of ImageNet-pretraining leakage. They propose FQAT, which adds a flatness-minimization objective to the QAT loss and adaptively freezes the quantization scale factors' original-QAT gradients when a gradient-disorder metric signals instability. In their experiments FQAT improves both in-distribution and OOD accuracy, with the larger gains on OOD, and at 8 bits it surpasses the full-precision model's OOD accuracy on DomainNet. The practically relevant claim: quantization can quietly undermine generalization, and a flatness-aware schedule is a workable remedy.

What carries the argument

The load-bearing mechanisms are two. First, a flatness-oriented QAT objective built from SAGM: the training loss is L(Q(θ;s);D) plus a perturbed loss L_p(Q(θ−α∇L(Q(θ;s);D);s);D), so the model must keep both low loss and a flat minimum. Second, the gradient-disorder metric δ, defined as the fraction of adjacent training steps in a K-step window where the sign of a scale factor's vanilla QAT gradient flips. A threshold r on δ drives an adaptive freezing algorithm: when δ<r, the vanilla-QAT gradient for that scale factor is frozen out of the update (only the flatness gradient acts); otherwise both gradients update. This keeps volatile layers alive and stops low-disorder layers from cancelling t

What would settle it

Run the same MoCoV2-pretrained ERM backbone through LSQ and EWGS while sweeping the scale-factor learning rate (say 1e-6 to 1e-3) and the number of training steps (say 5,000 to 50,000), recording OOD test accuracy on PACS; if any setting restores OOD accuracy to the full-precision model's level without a flatness objective, the claim that standard QAT degrades OOD generalization would be falsified. A complementary check is repeating the PACS Art 4-bit experiment with a different self-supervised pretraining (e.g., DINO) to see whether the drop reproduces.

Watch

Extended reading notes

Core claim

The paper's central discovery is that existing quantization-aware training methods induce a sharp loss landscape, and that this sharpness is the mechanism by which quantization degrades out-of-distribution generalization. Working with a MoCoV2-pretrained ResNet-50 fine-tuned by empirical risk minimization to avoid ImageNet pretraining leakage, the authors show that LSQ and EWGS quantization at 3-8 bits produces large OOD test-accuracy drops on PACS, OfficeHome, and DomainNet. They trace the problem to gradient conflict in the scale-factor updates: the vanilla QAT gradient and the flatness-oriented gradient often point in opposite directions, locking scale factors in a sub-optimal equilibrium

Load-bearing premise

The conclusion that QAT inherently degrades OOD accuracy rests on one fixed recipe — a MoCoV2-pretrained ResNet-50 tuned with ERM, then quantized at a single scale-factor learning rate (1e-5) and a fixed number of steps — and assumes the resulting drops are the fault of quantization rather than of those untuned QAT settings.

Editorial extensions

If this is right

  • Under the paper's setup, QAT should be treated as a deployment-relevant step: in-distribution accuracy alone overstates a quantized model's readiness for distribution shift.
  • The gradient-disorder freezing rule drops into existing learned-scale-factor QAT pipelines (demonstrated on LSQ and EWGS) and delivers consistent I.D and OOD gains without architectural changes.
  • Flatness objectives known to help full-precision domain generalization carry over to quantized training, but only after resolving the conflict with QAT gradients.
  • At 8-bit quantization on DomainNet, FQAT's OOD accuracy exceeds the full-precision ERM baseline, suggesting that at mild bit widths the flatness objective can more than compensate for quantization loss.
  • Gains are consistently larger on OOD than I.D (PACS 3-bit: +5.24% OOD vs +2.61% I.D over the direct-SAGM baseline), so the benefit is not merely recovering lost memorization.

Reading between the lines

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

  • The paper's fixed QAT schedule (scale-factor learning rate 1e-5, fixed step counts) leaves open that part of the reported drop is an artifact of undertuning; sweeping those hyperparameters and seeing whether the degradation persists would sharpen or weaken the claim.
  • Gradient-disorder freezing is a general recipe for multi-objective training where one objective's gradients are more volatile; it could be tested on full-precision domain generalization or on mixed-precision quantization with heterogeneous scale factors.
  • The leakage explanation generalizes as a warning: supervised ImageNet pretraining may mask robustness losses that self-supervised initializations reveal, so OOD evaluations of quantization should prefer leakage-free initializations or report both.
  • The loss-landscape plots show FQAT also stabilizes unfrozen layers (fewer gradient outliers), hinting that the freezing rule could combine well with weight averaging or other training stabilizers; the paper does not test that combination.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper studies whether quantization-aware training (QAT) degrades the out-of-distribution (OOD) generalization of full-precision vision models. It claims that existing QAT methods, such as LSQ and EWGS, significantly reduce OOD accuracy, and attributes this to the sharp loss landscape induced by quantization. The authors propose FQAT, which combines a SAGM-style flatness objective with a layer-wise, gradient-disorder-guided freezing mechanism for quantizer scale factors. Experiments on PACS, OfficeHome, and DomainNet compare LSQ, EWGS, LSQ+SAGM, EWGS+SAGM, and FQAT at several bit widths. The reported results show consistent gains for FQAT over the QAT baselines on both I.D validation and OOD test accuracy. The central empirical premise, however, is not yet established because the comparison lacks a full-precision control trained for the same additional steps, and the QAT baselines exhibit very large I.D accuracy drops that suggest undertuning.

Significance. If the central claim were well supported, this would be a useful first systematic study of OOD behavior in QAT and a practical contribution to deploying quantized models under distribution shift. The proposed freezing mechanism is novel and the authors provide code, loss-landscape visualizations, and a protocol that uses MoCoV2 pretraining to address ImageNet leakage concerns. The method's gains over LSQ+SAGM and EWGS+SAGM are consistent across most settings, which is a meaningful empirical result. However, the paper's main motivation—that QAT inherently harms OOD generalization—rests on a comparison with no full-precision fine-tuning control and with QAT baselines that lose a large amount of I.D accuracy. Until these confounds are addressed, the significance of the claimed phenomenon is uncertain, although the proposed method may still be practically useful.

major comments (4)
  1. [§5.1, Tables 1/2] The full-precision ERM baseline is trained for 5,000 steps on PACS/OfficeHome and 15,000 on DomainNet, while every QAT method is initialized from that model and trained for 20,000/50,000 additional steps. There is no full-precision control fine-tuned for the same additional steps under the same optimizer and schedule. The large OOD drops (e.g., PACS test 85.29 vs. LSQ 5-bit 57.38) could therefore reflect continued ERM fine-tuning on the source domains rather than an effect of quantization. Add a full-precision fine-tuning control and report its I.D/OOD accuracy; otherwise the Sec. 3 claim that 'quantization leads to OOD degradation' is not identified.
  2. [§5.1, Tables 1/2] The QAT baselines show surprisingly large I.D. drops: PACS LSQ 4-bit validation falls from 96.42 to 80.44, OfficeHome LSQ 3-bit from 78.44 to 59.58, and EWGS 3-bit from 78.44 to 60.32. A properly tuned QAT method should preserve I.D. accuracy much more closely; these numbers suggest the fixed scale-factor learning rate (1e-5) and/or the long training schedule are not well matched to LSQ/EWGS. DomainNet, where the I.D. drop is only about 1 point, also shows only a small OOD drop (about 1.8 points), which is consistent with this confound. Please tune the baselines and report whether the OOD degradation persists when I.D. accuracy is preserved.
  3. [§3, 'Conclusion' paragraph] The paper explains the discrepancy with QT-DoG by ImageNet pretraining leakage, but the supporting evidence is a single ablation: PACS Art, LSQ 4-bit, MoCoV2 (51.07) vs. ImageNet (76.51). This does not reproduce QT-DoG and does not rule out the alternative explanation that the QAT setup itself is undertuned. Please provide a more systematic comparison across domains and bit widths, and include a full-precision ImageNet-pretrained control for the same fine-tuning schedule.
  4. [Appendix B, Algorithm 1] Lines 10–18 evaluate the gradient disorder of gva for every scale factor every K steps, but lines 4–8 specify that for a frozen layer only gflat is used to update s_i, so gva is not computed during those intervals. The algorithm is underspecified about how delta_{t,s_i} is obtained for a layer whose gva has been frozen for several periods—whether gva is still computed for monitoring or stale histories are used. This ambiguity affects reproducibility and should be clarified.
minor comments (6)
  1. [Fig. 1 caption] 'quantified model' should be 'quantized model'.
  2. [§5.1] The sentence 'as 3bit quantization yields minimal conflicts (Figure 5)' seems to contradict applying LSQ/EWGS at 4/5/8 bit on DomainNet. Clarify whether 3-bit was excluded because the flatness conflict is minimal.
  3. [Definition 4.1] The notation sgn(G(i)_1) is confusing; use sgn(g_i) and sgn(g_{i+1}) explicitly.
  4. [Abstract / throughout] Use a consistent abbreviation for in-distribution (ID or I.D.) throughout.
  5. [Eq. (3)] It is not immediately clear how s is optimized with two gradients and how the perturbation is applied to weights but not scale factors. This is only explained in Appendix A; add a sentence in Sec. 4.1.
  6. [Table 4 caption] State explicitly that the ablation is run on a single seed (seed 0), since the main tables use two seeds.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: FQAT's improvements are evaluated on held-out OOD domains, and its premises rest on external citations and in-paper ablations rather than on fitting the target outputs.

full rationale

The paper's central claim is that QAT degrades OOD generalization and that the proposed FQAT method mitigates this by flattening the loss landscape and freezing unstable scale-factor gradients. Walking the derivation chain, I find no step that reduces by construction to its own inputs. The flatness-OOD link is imported from external prior work (SAGM, SWAD, etc.), not from a self-citation chain that is itself unverified. The claim that QAT sharpens the loss landscape is cited to an external paper (SAQ) and is further supported by the paper's own loss-landscape visualizations. The choice of MoCo-v2 over ImageNet pretraining is justified by a published external result and is additionally supported by an in-paper ablation (PACS Art, 4-bit LSQ: 51.07 vs 76.51). FQAT's freezing threshold and freeze steps are selected via grid search on validation-style single-domain observations, not on the OOD test domains used for evaluation; the method's gains are then reported on held-out test domains, so the central results are not fitted quantities renamed as predictions. The gradient-disorder metric determines freezing from training dynamics, not from test labels. There is a legitimate experimental-design concern that the QAT baselines may be undertuned (fixed scale-factor LR, longer training without a full-precision control), but that is a correctness or confound issue, not circularity: the paper does not define its OOD prediction in terms of the fitted QAT hyperparameters. Similarly, the self-citations to the authors' prior work (e.g., [48] on ImageNet leakage, [40] on freezing) are external published results with independent evidence and do not constitute load-bearing self-citation. Overall, the derivation is self-contained against external benchmarks, and no circular step can be exhibited.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a standard empirical domain-generalization setup plus a new heuristic freezing rule. The main free parameters are the freezing threshold and interval, both tuned per dataset and bit-width. There are no invented physical entities. The assumptions about flatness and OOD generalization are inherited from prior work and are the main theoretical load-bearing elements.

free parameters (5)
  • gradient disorder threshold r = 0.28-0.32 for PACS, 0.20-0.25 or 0.30-0.35 for DomainNet, depending on bit-width
    Hand-tuned criterion for freezing the vanilla QAT gradient; selected via grid search per dataset and bit-width.
  • freeze interval K = 300-400 steps for PACS, 100-200 for 3-bit, 3000-4000 for DomainNet 4/5-bit, 1500-2000 for 8-bit
    Number of steps over which gradient disorder is measured and freezing is applied; grid-searched per setting.
  • SAGM alpha = 0.001 for PACS and OfficeHome, 0.0005 for DomainNet
    Step size for the perturbed-weight update in the flatness objective, taken from SAGM but tuned for the QAT setting.
  • SAM rho = 0.05
    Perturbation radius in the flatness objective, fixed following SAM.
  • scale factor learning rate = 1e-5
    Learning rate for all learned quantization scale factors; chosen by hand and not part of the main grid search.
assumptions (4)
  • domain assumption Flatter loss landscapes improve OOD generalization.
    Taken from SAM and SAGM literature; this is the theoretical motivation for adding a flatness objective.
  • domain assumption Quantization sharpens the loss landscape, which causes OOD degradation.
    Cites SAQ [27] and visualizes the loss landscape in Figure 1, but the causal link is not quantified.
  • ad hoc to paper The gradient disorder metric identifies layers where freezing the QAT gradient is beneficial.
    Assumption 4.2 states that scale factors with low gva gradient disorder hinder their own flatness training; this is a heuristic validated only through the paper's own experiments.
  • domain assumption MoCoV2-pretrained ResNet50 is the correct evaluation base because ImageNet pretraining leaks OOD information.
    Follows Yu et al. [48]; the paper's single-domain ablation supports this, but it is not proven exhaustively.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantization Meets OOD: Generalizable Quantization-aware Training from a Flatness Perspective." pith.science (2026). https://pith.science/paper/3Y3KPAK7

@misc{pith2026250900859,
  author       = {Pith},
  title        = {Pith review of: Quantization Meets OOD: Generalizable Quantization-aware Training from a Flatness Perspective},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3Y3KPAK7}},
  note         = {Machine review of arXiv:2509.00859}
}
read the original abstract

Current quantization-aware training (QAT) methods primarily focus on enhancing the performance of quantized models on in-distribution (I.D) data, while overlooking the potential performance degradation on out-of-distribution (OOD) data. In this paper, we first substantiate this problem through rigorous experiment, showing that QAT can lead to a significant OOD generalization performance degradation. Further, we find the contradiction between the perspective that flatness of loss landscape gives rise to superior OOD generalization and the phenomenon that QAT lead to a sharp loss landscape, can cause the above problem. Therefore, we propose a flatness-oriented QAT method, FQAT, to achieve generalizable QAT. Specifically, i) FQAT introduces a layer-wise freezing mechanism to mitigate the gradient conflict issue between dual optimization objectives (i.e., vanilla QAT and flatness). ii) FQAT proposes an disorder-guided adaptive freezing algorithm to dynamically determines which layers to freeze at each training step, effectively addressing the challenges caused by interference between layers. A gradient disorder metric is designed to help the algorithm identify unstable layers during training. Extensive experiments on influential OOD benchmark demonstrate the superiority of our method over state-of-the-art baselines under both I.D and OOD image classification tasks.

Figures

Figures reproduced from arXiv: 2509.00859 by the authors.

Figure 1
Figure 1. Visualization of loss landscapes of full precision [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Results of cumulative gradients every 350 steps in [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Results of the 4-bit quantization test on the PACS [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Results of cumulative gradients every 2111 steps [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: 4-bit LSQ PACS (seed 0, 23): Val acc (I.D mean) and [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Visualization of the loss landscape across different domains (PACS 4bit), where the categories from left to right are Art, Cartoon, Photo, and Sketch. Blue indicates low loss values, while red represents high loss values. The top row displays the results of LSQ + SAGM,…
Figure 8
Figure 8. Figure 8: Visualization of the loss landscape of DomainNet Clipart & Infograph domains (4bit). The top row displays the results [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

60 extracted references · 44 canonical work pages

  1. [1]

    Yogesh Balaji, Swami Sankaranarayanan, and Rama Chellappa. 2018. Metareg: Towards domain generalization using meta-regularization. Advances in neural information processing systems 31 (2018)

  2. [2]

    Yoshua Bengio, Nicholas Léonard, and Aaron Courville. 2013. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432 (2013)

  3. [3]

    Fabio M Carlucci, Antonio D’Innocente, Silvia Bucci, Barbara Caputo, and Tatiana Tommasi. 2019. Domain generalization by solving jigsaw puzzles. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 2229–2238

  4. [4]

    Junbum Cha, Sanghyuk Chun, Kyungjae Lee, Han-Cheol Cho, Seunghyun Park, Yunsung Lee, and Sungrae Park. 2021. Swad: Domain generalization by seeking flat minima. Advances in Neural Information Processing Systems 34 (2021), 22405– 22418

  5. [5]

    Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. 2020. Improved Baselines with Momentum Contrastive Learning. arXiv:2003.04297 [cs.CV] https://arxiv. org/abs/2003.04297

  6. [6]

    Jungwook Choi, Zhuo Wang, Swagath Venkataramani, Pierce I-Jen Chuang, Vijayalakshmi Srinivasan, and Kailash Gopalakrishnan. 2018. Pact: Parameterized clipping activation for quantized neural networks.arXiv preprint arXiv:1805.06085 (2018)

  7. [7]

    Qi Dou, Daniel Coelho de Castro, Konstantinos Kamnitsas, and Ben Glocker

  8. [8]

    Steven K Esser, Jeffrey L McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S Modha. 2019. Learned step size quantization. arXiv preprint arXiv:1902.08153 (2019)

Show all 60 references
  1. [9]

    Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. 2020. Sharpness-aware minimization for efficiently improving generalization. arXiv preprint arXiv:2010.01412 (2020)

  2. [10]

    Ishaan Gulrajani and David Lopez-Paz. 2020. In search of lost domain generaliza- tion. arXiv preprint arXiv:2007.01434 (2020)

  3. [11]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition . 770–778

  4. [12]

    Qiang Hu, Yuejun Guo, Maxime Cordy, Xiaofei Xie, Wei Ma, Mike Papadakis, and Yves Le Traon. 2022. Characterizing and understanding the behavior of quantized models for reliable deployment. arXiv preprint arXiv:2204.04220 (2022)

  5. [13]

    Zeyi Huang, Haohan Wang, Eric P Xing, and Dong Huang. 2020. Self-challenging improves cross-domain generalization. In Computer vision–ECCV 2020: 16th Euro- pean conference, Glasgow, UK, August 23–28, 2020, proceedings, part II 16 . Springer, 124–140

  6. [14]

    Itay Hubara, Yury Nahshan, Yair Hanani, Ron Banner, and Daniel Soudry. 2021. Accurate post training quantization with small calibration sets. In International Conference on Machine Learning . PMLR, 4466–4475

  7. [15]

    Brody Huval, Tao Wang, Sameep Tandon, Jeff Kiske, Will Song, Joel Pazhayam- pallil, Mykhaylo Andriluka, Pranav Rajpurkar, Toki Migimatsu, Royce Cheng-Yue, et al. 2015. An empirical evaluation of deep learning on highway driving. arXiv preprint arXiv:1504.01716 (2015)

  8. [16]

    Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and An- drew Gordon Wilson. 2018. Averaging weights leads to wider optima and better generalization. arXiv preprint arXiv:1803.05407 (2018)

  9. [17]

    Saqib Javed, Hieu Le, and Mathieu Salzmann. 2024. QT-DoG: Quantization-aware Training for Domain Generalization. arXiv preprint arXiv:2410.06020 (2024)

  10. [18]

    Norman P Jouppi, Cliff Young, Nishant Patil, David Patterson, Gaurav Agrawal, Raminder Bajwa, Sarah Bates, Suresh Bhatia, Nan Boden, Al Borchers, et al. 2017. In-datacenter performance analysis of a tensor processing unit. In Proceedings of the 44th annual international sympos...

  11. [19]

    Sangil Jung, Changyong Son, Seohyung Lee, Jinwoo Son, Jae-Joon Han, Youngjun Kwak, Sung Ju Hwang, and Changkyu Choi. 2019. Learning to quantize deep networks by optimizing quantization intervals with task loss. In Proceedings of the IEEE/CVF conference on computer vision and p...

  12. [20]

    Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas Phillips, Irena Gao, et al. 2021. Wilds: A benchmark of in-the-wild distribution shifts. In International conference on machine learn...

  13. [21]

    Junghyup Lee, Dohyung Kim, and Bumsub Ham. 2021. Network quantization with element-wise gradient scaling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 6448–6457

  14. [22]

    Jesse Levinson, Jake Askeland, Jan Becker, Jennifer Dolson, David Held, Soeren Kammel, J Zico Kolter, Dirk Langer, Oliver Pink, Vaughan Pratt, et al. 2011. To- wards fully autonomous driving: Systems and algorithms. In 2011 IEEE intelligent vehicles symposium (IV). IEEE, 163–168

  15. [23]

    Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy Hospedales. 2018. Learning to generalize: Meta-learning for domain generalization. In Proceedings of the AAAI conference on artificial intelligence , Vol. 32

  16. [24]

    Hospedales

    Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M. Hospedales. 2017. Deeper, Broader and Artier Domain Generalization. InProceedings of the IEEE International Conference on Computer Vision (ICCV)

  17. [25]

    Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. 2018. Visualizing the loss landscape of neural nets. Advances in neural information processing systems 31 (2018)

  18. [26]

    Ya Li, Mingming Gong, Xinmei Tian, Tongliang Liu, and Dacheng Tao. 2018. Domain generalization via conditional invariant representations. In Proceedings of the AAAI conference on artificial intelligence , Vol. 32

  19. [27]

    Jing Liu, Jianfei Cai, and Bohan Zhuang. 2021. Sharpness-aware quantization for deep neural networks. arXiv preprint arXiv:2111.12273 (2021)

  20. [28]

    Jiashuo Liu, Zheyan Shen, Yue He, Xingxuan Zhang, Renzhe Xu, Han Yu, and Peng Cui. 2021. Towards out-of-distribution generalization: A survey. arXiv preprint arXiv:2108.13624 (2021)

  21. [29]

    Shih-Yang Liu, Zechun Liu, and Kwang-Ting Cheng. 2023. Oscillation-free quan- tization for low-bit vision transformers. In International Conference on Machine Learning. PMLR, 21813–21824

  22. [30]

    Yong Liu, Siqi Mai, Xiangning Chen, Cho-Jui Hsieh, and Yang You. 2022. To- wards efficient and scalable sharpness-aware minimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 12360–12370

  23. [31]

    Fangrui Lv, Jian Liang, Shuang Li, Bin Zang, Chi Harold Liu, Ziteng Wang, and Di Liu. 2022. Causality inspired representation learning for domain generalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 8046–8056

  24. [32]

    Divyat Mahajan, Shruti Tople, and Amit Sharma. 2021. Domain generalization using causal matching. In International conference on machine learning . PMLR, 7313–7324

  25. [33]

    Krikamol Muandet, David Balduzzi, and Bernhard Schölkopf. 2013. Domain generalization via invariant feature representation. In International conference on machine learning. PMLR, 10–18

  26. [34]

    Markus Nagel, Rana Ali Amjad, Mart Van Baalen, Christos Louizos, and Tijmen Blankevoort. 2020. Up or Down? Adaptive Rounding for Post-Training Quanti- zation. In Proceedings of the 37th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol...

  27. [35]

    Markus Nagel, Marios Fournarakis, Rana Ali Amjad, Yelysei Bondarenko, Mart van Baalen, and Tijmen Blankevoort. 2021. A White Paper on Neural Network Quantization. arXiv:2106.08295 [cs.LG] https://arxiv.org/abs/2106.08295

  28. [36]

    Markus Nagel, Marios Fournarakis, Yelysei Bondarenko, and Tijmen Blankevoort

  29. [37]

    Xingchao Peng, Zijun Huang, Ximeng Sun, and Kate Saenko. 2019. Domain agnostic learning with disentangled representations. In International conference on machine learning. PMLR, 5102–5112

  30. [38]

    Jiantao Qiu, Jie Wang, Song Yao, Kaiyuan Guo, Boxun Li, Erjin Zhou, Jincheng Yu, Tianqi Tang, Ningyi Xu, Sen Song, et al. 2016. Going deeper with embedded FPGA platform for convolutional neural network. In Proceedings of the 2016 ACM/SIGDA international symposium on field-prog...

  31. [39]

    Juncheol Shin, Junhyuk So, Sein Park, Seungyeop Kang, Sungjoo Yoo, and Eun- hyeok Park. 2023. NIPQ: Noise proxy-based integrated pseudo-quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion. 3852–3861

  32. [40]

    Chen Tang, Yuan Meng, Jiacheng Jiang, Shuzhao Xie, Rongwei Lu, Xinzhu Ma, Zhi Wang, and Wenwu Zhu. 2024. Retraining-free model quantization via one- shot weight-coupling learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 15855–15865

  33. [41]

    Chen Tang, Kai Ouyang, Zhi Wang, Yifei Zhu, Wen Ji, Yaowei Wang, and Wenwu Zhu. 2022. Mixed-precision neural network quantization via learned layer-wise importance. In European Conference on Computer Vision . Springer, 259–275

  34. [42]

    Vladimir Vapnik. 1991. Principles of risk minimization for learning theory. Advances in neural information processing systems 4 (1991)

  35. [43]

    Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. 2017. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition . 5018–5027. MM ’25, October 27–31, 2025, Dublin...

  36. [44]

    Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and S Yu Philip. 2022. Generalizing to unseen domains: A survey on domain generalization. IEEE transactions on knowledge and data engineering 35, 8 (2022), 8052–8072

  37. [45]

    Pengfei Wang, Zhaoxiang Zhang, Zhen Lei, and Lei Zhang. 2023. Sharpness-aware gradient matching for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 3769–3778

  38. [46]

    Jiaming Yang, Chenwei Tang, Caiyang Yu, and Jiancheng Lv. 2024. GWQ: Group- Wise Quantization Framework for Neural Networks. In Asian Conference on Machine Learning. PMLR, 1526–1541

  39. [47]

    Han Yu, Jiashuo Liu, Xingxuan Zhang, Jiayun Wu, and Peng Cui. 2024. A survey on evaluation of out-of-distribution generalization. arXiv preprint arXiv:2403.01874 (2024)

  40. [48]

    Han Yu, Xingxuan Zhang, Renzhe Xu, Jiashuo Liu, Yue He, and Peng Cui. 2024. Rethinking the evaluation protocol of domain generalization. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 21897–21908

  41. [49]

    Hanlin Zhang, Yi-Fan Zhang, Weiyang Liu, Adrian Weller, Bernhard Schölkopf, and Eric P Xing. 2022. Towards principled disentanglement for domain general- ization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 8024–8034

  42. [50]

    Xingxuan Zhang, Renzhe Xu, Han Yu, Yancheng Dong, Pengfei Tian, and Peng Cui. 2023. Flatness-aware minimization for domain generalization. InProceedings of the IEEE/CVF International Conference on Computer Vision . 5189–5202

  43. [51]

    Xingxuan Zhang, Renzhe Xu, Han Yu, Hao Zou, and Peng Cui. 2023. Gradient norm aware minimization seeks first-order flatness and improves generaliza- tion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 20247–20257

  44. [52]

    Shanshan Zhao, Mingming Gong, Tongliang Liu, Huan Fu, and Dacheng Tao

  45. [53]

    Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. 2022. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 45, 4 (2022), 4396–4415

  46. [54]

    Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. 2021. Domain generaliza- tion with mixstyle. arXiv preprint arXiv:2104.02008 (2021)

  47. [55]

    Shuchang Zhou, Yuxin Wu, Zekun Ni, Xinyu Zhou, He Wen, and Yuheng Zou

  48. [56]

    Juntang Zhuang, Boqing Gong, Liangzhe Yuan, Yin Cui, Hartwig Adam, Nicha C Dvornek, James s Duncan, Ting Liu, et al . 2022. Surrogate Gap Minimization Improves Sharpness-Aware Training. In International Conference on Learning Representations. Quantization Meets OOD: Generaliza...

  49. [2016]

    arXiv preprint arXiv:1606.06160 (2016)

    Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint arXiv:1606.06160 (2016)

  50. [2019]

    Advances in neural information processing systems 32 (2019)

    Domain generalization via model-agnostic learning of semantic features. Advances in neural information processing systems 32 (2019)

  51. [2020]

    Advances in neural information processing systems 33 (2020), 16096–16107

    Domain generalization via entropy regularization. Advances in neural information processing systems 33 (2020), 16096–16107

  52. [2022]

    In International Conference on Machine Learning

    Overcoming oscillations in quantization-aware training. In International Conference on Machine Learning . PMLR, 16318–16330

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.