Pith. sign in

REVIEW 4 major objections 5 minor 35 references

Efficient Tuning Before Low-Bit Post-Training Quantization for Stochastic Gradient Descent-optimized Models

T0 review · 4 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read Low-bit post-training quantization accuracy is capped by the full-precision model's tolerance to quantization error, and ETBQ shows that tolerance can be improved beforehand, yielding 2.14-point top-1 gains on Tiny-ImageNet and 5.80-point m

desk verdict A serious, honestly written empirical paper that shows an upstream fine-tuning stage can improve low-bit PTQ; the quantizer-alignment mechanism is plausible but untested—worth peer review with requested controls. read the letter →

arxiv 2607.11359 v2 pith:A6MKNGWG submitted 2026-07-13 cs.CV

Peng Xia , Junbiao Pang , Muhammad Ayub Sabir This is my paper · ORCID
classification cs.CV
keywords post-trainingquantizationlow-bitquantization-awaretuningerrortolerancelosslandscapeflatnessstochasticweightaveragingnoiseactivation
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 argues that low-bit post-training quantization (PTQ) accuracy is limited not only by how well a PTQ backend reconstructs the full-precision model, but also by how sensitive the full-precision model itself is to quantization errors. To address this, it introduces ETBQ, a full-precision tuning stage placed before PTQ: the model is optimized under weight and activation quantization-error perturbations sampled from statistics estimated on a calibration set, then averaged with stochastic weight averaging. The resulting full-precision model is used by an unchanged reconstruction-based PTQ backend, and the paper reports consistent gains across CIFAR-100, Tiny-ImageNet, ImageNet, and Cityscapes, with the largest improvements at W2A4. If the claim holds, it means quantization tolerance can be built into the full-precision model upstream rather than repaired only after quantization.

What carries the argument

The core mechanism is calibration-driven perturbation conditioning. WQN simulates the target channel-wise weight quantizer, estimates per-channel Gaussian error statistics, and forms a differenced perturbation P_t = δ_t − δ_{t−1} so that non-zero-mean quantization error does not accumulate in the FP weights; the task loss is evaluated at the perturbed weights while the stored weights remain clean. AQN estimates per-tensor Gaussian statistics for activation quantization error and samples an error tensor, then applies a salience-aware Bernoulli mask so that only channels with larger spatial absolute responses receive the perturbation. SWA then averages the late-training weights, and the paper

What would settle it

Run an otherwise identical full-precision fine-tuning control that uses the same epochs, data, and compute but omits WQN and AQN perturbations; if that control matches ETBQ's W2A4 accuracy, the perturbation mechanism is not the cause. Additionally, at W2A2, compare the actual per-layer weight and activation quantization errors against the Gaussian model; large tail deviation would predict the small gains observed.

Watch

Extended reading notes

Core claim

The central claim is that a full-precision model can be pre-conditioned, before any PTQ backend runs, to be more tolerant of low-bit quantization error. ETBQ does this by injecting two calibration-estimated perturbation types during an SGD fine-tuning stage: Weight Quantization Noise, which samples from the channel-wise weight-quantization error distribution and uses temporal differencing to control drift, and Activation Quantization Noise, which samples from per-tensor activation-quantization error and applies a salience-aware stochastic mask. Stochastic weight averaging is then used to stabilize the trajectory, and BatchNorm statistics are recomputed. The output is a full-precision model t

Load-bearing premise

The central claim depends on the assumption that Gaussian noise sampled from calibration-set error statistics faithfully represents the errors the downstream quantizer will actually introduce at low bit-widths, and that this specific conditioning—not ordinary fine-tuning—is what improves downstream PTQ accuracy.

Editorial extensions

If this is right

  • If the claim is correct, PTQ users can improve low-bit accuracy without modifying the quantizer or training a fake-quantized model, simply by pre-tuning the full-precision model.
  • The benefit is strongest at W2A4, where quantization error is severe but the model has not collapsed; gains at W2A2 are smaller, suggesting a robustness boundary at extremely low bit-widths.
  • ETBQ sits between calibration-only PTQ and full quantization-aware training in cost, using a full-data fine-tuning stage of about 80 epochs on ImageNet while leaving quantizer optimization to the downstream PTQ backend.
  • The preconditioned model is backend-agnostic: the paper reports gains with two reconstruction-based PTQ backends across classification and semantic segmentation, indicating the improvement is not tied to one calibration method.

Reading between the lines

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

  • Editorial inference: If flatness of the loss landscape is the operative cause, a sharpness-penalizing fine-tuning scheme that does not require calibration-set error statistics should reproduce much of ETBQ's gain; this is directly testable.
  • Editorial inference: The Gaussian surrogate for quantization error is the method's fragile link, and the small W2A2 gains give a natural place to test whether heavy-tailed or structured quantizer error violates the model's assumptions.
  • Editorial inference: Because ETBQ also improves FP32 accuracy slightly, part of the gain may be regularization rather than quantization-specific conditioning; a control that fine-tunes with the same compute but without WQN/AQN would separate these contributions.
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 / 5 minor

Summary. ETBQ proposes an upstream full-precision (FP) pre-conditioning stage for SGD-trained models before low-bit post-training quantization (PTQ). During tuning, it injects Weight Quantization Noise (WQN), modeled as channel-wise Gaussians with temporal differencing (Eqs. 5–9), and Activation Quantization Noise (AQN), modeled as per-tensor Gaussians with salience-aware stochastic masking (Eqs. 10–15), followed by Stochastic Weight Averaging (SWA). The resulting FP model is then fed to standard PTQ backends such as QDrop and QEP. Experiments on CIFAR-100, Tiny-ImageNet, ImageNet, and Cityscapes report consistent improvements, e.g., W2A4 top-1 on Tiny-ImageNet ResNet-18 rises from 63.71 to 65.85, and Cityscapes U-Net mIoU from 57.72 to 63.52. The paper argues that ETBQ moves the FP model to a flatter loss basin, making it more tolerant to quantization errors, and contrasts the method with calibration-only PTQ and full QAT.

Significance. If the claims hold, ETBQ is a simple, backend-agnostic way to improve low-bit PTQ by improving the FP starting point, with compute positioned between calibration-only PTQ and full QAT (Table VII). The paper's strengths are its broad empirical coverage (five architectures, four datasets, two PTQ backends), released code, and honest reporting of component ablations and limitations, including the explicit statement that the flatness evidence is correlational and the acknowledgment that the Gaussian surrogate degrades at W2A2. The core empirical result — that the integrated ETBQ pipeline improves downstream PTQ across diverse settings — appears internally consistent. The main gap is that the mechanism claim (quantization-aligned perturbations, not generic fine-tuning/noise, drive the gains) is not isolated experimentally: there is no plain-fine-tuning control and no isotropic-noise control. The central empirical contribution is credible, but the attribution of the effect to quantization-aligned structure is under-tested.

major comments (4)
  1. [Section V-B, Table II] The ablation table has no plain-fine-tuning control. Table II includes WQN-only, AQN-only, SWA-only, and combinations, but no row with the same 120-epoch SGD schedule, cosine annealing, label smoothing, and SWA but without any WQN/AQN noise injection. Since the full ETBQ gain over direct QDrop is 1.59 pp (78.06 vs 76.47), and SWA alone already gives 77.22, the incremental contribution of quantization-aligned perturbations could be much smaller than the contribution of extra training plus SWA. Add a control row: same schedule, same optimizer, same SWA, no noise injection. Without it, the paper's central claim that pre-conditioning with quantization-error samples is what improves PTQ is not established.
  2. [Section IV-A, Appendix C-A] The mechanism claim distinguishes WQN from 'generic random perturbation' and 'generic anti-correlated injection,' but the experiments do not include an isotropic-noise control. Table IX compares naive additive vs differential injection, but both use quantization-aligned errors. A decisive control is to inject isotropic Gaussian noise with the same per-channel variance and the same temporal differencing, but with directions not aligned to the target quantizer's error distribution. If this control reproduces the W2A4 gains, the quantization-aligned component of WQN is not necessary. This is load-bearing because the paper's stated contribution is alignment with the target quantizer's error distribution, not merely noise-induced flatness.
  3. [Table VIII, Eq. (13)] The hyperparameter rho (target AQN mask density) is never reported, and lambda_max is tuned per dataset without sensitivity analysis. Table VIII lists lambda_max = 1.0, 0.8, 0.05, 0.8 for CIFAR-100, Tiny-ImageNet, ImageNet, and Cityscapes, but no ablation shows how sensitive results are to this choice. Since Eq. (13) uses rho to control how many channels receive activation perturbation, omitting its value makes the method hard to reproduce and leaves open the possibility that results depend on finely tuned noise intensity. Report rho for every dataset and provide a small sensitivity study (e.g., lambda_max in {0.2, 0.5, 1.0} for ResNet-18/CIFAR-100 W2A4).
  4. [Section VI, Appendix A-A] The paper acknowledges that the Gaussian surrogate becomes inaccurate under 2-bit quantization (NRMSE_QQ increases to about 0.10) and that W2A2 gains are smaller. This is not fatal, but it weakens the statement in Section IV-A that WQN/AQN model 'the genuine quantization error' rather than a generic perturbation. In the very regime emphasized in the title (2-bit), the moment-matched Gaussian is a loose approximation. Either soften the modeling claim, or add a comparison where the perturbation is sampled directly from the empirical error distribution on the calibration set for W2A2, to test whether the Gaussian approximation, rather than the alignment principle, limits the method.
minor comments (5)
  1. [Section V-B, Table II text] Typo: 'WQA+AQN+SWA' should be 'WQN+AQN+SWA'.
  2. [Section IV-A] The paragraph beginning 'Transferred activation quantization error overwhelms the weight quantization one.' is duplicated verbatim within the same section; remove the repetition.
  3. [Table VIII] The initial learning rate entry for Tiny-ImageNet reads '0.0012×10^{-5}', which appears to be a formatting corruption. Please clarify the intended value.
  4. [Fig. 1, Section I] The Hessian spectral density comparison is used as motivational evidence for flatness, but the text does not specify how the Hessian is estimated (e.g., full Hessian, Hutchinson trace estimate, top-k power iteration). Since the flatness claim is central to the narrative, add one or two sentences on the estimation procedure (in the main text or appendix).
  5. [Algorithm 1, line 4] The algorithm says 'Refresh WQN and AQN statistics on D_c' every epoch, but Section IV-B says AQN statistics are updated via EMA with beta=0.9. Clarify whether the refresh is a full recomputation or an EMA update, and at what frequency (per epoch or per iteration).

Circularity Check

0 steps flagged · score 2.0 of 10

No construction-level circularity: ETBQ's perturbations are estimated from calibration-set quantization errors, and the reported PTQ accuracy is measured post-hoc, not fitted. Minor self-citation and a missing generic-noise control weaken the mechanism claim but do not make any equation reduce to its inputs.

full rationale

ETBQ does not reduce to its own inputs. WQN/AQN statistics (Eqs. 5 and 11) are estimated from calibration-set quantization errors of the target quantizer, but the reported predictions are task accuracy/mIoU of the downstream QDrop/QEP quantized model on held-out data. The training loss (Eq. 16) is the task loss under perturbed weights/activations, not the reported metric, so no fitted parameter is renamed as a prediction. The Hessian-flatness evidence is explicitly presented as correlation ('These empirical findings establish a correlation, rather than a strict causal proof') and is measured after the fact, not used as the training target. The only attribution concern is reference [24], a self-citation by a coauthor, used as provenance for AQN's quantization-aligned rather than generic noise design; this is non-load-bearing because the method is specified by this paper's own equations and component ablations, and the headline gains would stand even if generic noise were equally effective. The open limitation that the Gaussian surrogate degrades at W2A2, and the absence of an isotropic-noise control, are experimental-control gaps rather than construction-level circularity: they weaken the mechanism claim but do not make any equation equal to another by construction. Overall score reflects the minor self-citation and no construction-level circularity.

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

The ledger shows that ETBQ introduces no free physical constants, but its effectiveness depends on several per-dataset hyperparameters (notably lambda_max and the unreported rho), a Gaussian surrogate, and an unproven flatness mechanism. No new physical entities are postulated.

free parameters (5)
  • lambda_max (maximum noise intensity) = 1.0 (CIFAR-100), 0.8 (Tiny-ImageNet), 0.05 (ImageNet), 0.8 (Cityscapes)
    Controls WQN/AQN perturbation strength in Eq. (17); varies by orders of magnitude across datasets, indicating hand-tuning rather than a derived setting.
  • rho (AQN mask density)
    Defined in Eq. (13) as C*rho expected selected channels; never reported in Table VIII or in the ablations, so a central architectural hyperparameter is unspecified.
  • SWA schedule = final 60 epochs on CIFAR/Tiny-ImageNet; 100 SWA epochs after 300 clean on Cityscapes; full 80-epoch ImageNet run
    SWA start and length differ by dataset and directly determine the flatness/consolidation effect claimed by ETBQ.
  • AQN EMA momentum beta = 0.9
    Used in Eq. (21) to smooth activation-error statistics; fixed by hand across all experiments.
  • Warmup epochs E_warmup = 20 / 20 / 3 / 10
    Schedule for the lambda_e ramp in Eq. (17); dataset-specific and part of the method's sensitivity.
assumptions (4)
  • domain assumption Quantization errors of weights and activations are adequately approximated by Gaussian distributions.
    Eqs. (5) and (11) and Appendix A-A; the authors report r_QQ > 0.99 but NRMSE worsens to about 0.10 at 2-bit, so the approximation is a first-order surrogate and less accurate at very low bit-widths.
  • domain assumption Error statistics estimated on the calibration subset remain representative during pre-conditioning.
    WQN/AQN statistics are recomputed from the calibration subset D_c (Algorithm 1, line 5); if the subset represents the training distribution poorly, the injected noise is misaligned with the downstream PTQ errors.
  • domain assumption Flatter loss basins imply better quantization tolerance.
    Hessian evidence in Fig. 1 is correlational; the paper explicitly disclaims causal proof, yet the method's design uses flatness as the mechanism for robustness.
  • domain assumption The method is only assumed to work for SGD-optimized models.
    Title, abstract, and Section VI state that Adam-trained models are not covered; transfer of the pre-conditioning idea to other optimizers is untested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficient Tuning Before Low-Bit Post-Training Quantization for Stochastic Gradient Descent-optimized Models." pith.science (2026). https://pith.science/paper/A6MKNGWG

@misc{pith2026260711359,
  author       = {Pith},
  title        = {Pith review of: Efficient Tuning Before Low-Bit Post-Training Quantization for Stochastic Gradient Descent-optimized Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A6MKNGWG}},
  note         = {Machine review of arXiv:2607.11359}
}
read the original abstract

Post-training quantization (PTQ) compresses deep neural networks for deployment under limited memory and computational budgets. However, low-bit (i.e., 2-bit or 4-bit) PTQ often suffers from substantial performance degradation. Most existing PTQ methods operate on an unconstrained full-precision (FP) model and primarily address quantization errors through post-hoc reconstruction. We argue that low-bit PTQ accuracy is limited not only by post-quantization error minimization, but also by the quantization-error tolerance of a FP model itself. In this paper, we propose Efficient Tuning Before Quantization (ETBQ), a pre-conditioning tuning stage for Stochastic Gradient Descent (SGD)-optimized models before PTQ. During tuning, the FP model is optimized under perturbations sampled from the error distributions of weight and activation quantization, guiding the model toward a loss-landscape region that is less sensitive to the subsequent PTQ. Unlike QAT, ETBQ does not train a fake-quantized deployment model, which is computationally and memory intensive. Instead, ETBQ outputs a FP model that can be used by any PTQ backend. Experiments on CIFAR-100, Tiny-ImageNet, ImageNet, and Cityscapes provide consistent evidence that ETBQ improves low-bit PTQ across diverse tasks. Under W2A4 settings, e.g., ETBQ improves over naive PTQ by 2.14\% top-1 accuracy on Tiny-ImageNet and by 5.80\% mIoU on Cityscapes. Code is available at https://github.com/xpxpxp2001xpxpxp/ETBQ.

Figures

Figures reproduced from arXiv: 2607.11359 by the authors.

Figure 1
Figure 1. Comparison between Hessian spectral densities of ResNet-18 on [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Comparisons between weight and transferred activation quantization [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Q-Q diagnostics based on the digitized transferred-error histogram [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: ETBQ inserts a pre-conditioning stage between FP training and PTQ. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Layer-wise quantization error reduction ratio of ETBQ on ResNet [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Representative weight and quantization-error distributions from [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Gaussian Q-Q diagnostics for representative channel-wise weight [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 5 linked inside Pith

  1. [1]

    Brecq: Pushing the limit of post-training quantization by block reconstruction,

    Y . Li, R. Gong, X. Tan, Y . Yang, P. Hu, Q. Zhang, F. Yu, W. Wang, and S. Gu, “Brecq: Pushing the limit of post-training quantization by block reconstruction,” inInternational Conference on Learning Representations (ICLR), 2021

    2021

  2. [2]

    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

    2020

  3. [3]

    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

  4. [4]

    Qdrop: Randomly dropping quantization for extremely low-bit post-training quantization,

    X. Wei, R. Gong, Y . Li, X. Liu, and F. Yu, “Qdrop: Randomly dropping quantization for extremely low-bit post-training quantization,” inInternational Conference on Learning Representations (ICLR), 2022

    2022

  5. [5]

    Flatquant: Flatness matters for llm quantization,

    R. Liuet al., “Flatquant: Flatness matters for llm quantization,” inICLR, 2025

    2025

  6. [6]

    Lsq+: Improving low-bit quantization through learnable offsets and better initialization,

    Y . Bhalgat, J. Lee, M. Nagel, T. Blankevoort, and N. Kwak, “Lsq+: Improving low-bit quantization through learnable offsets and better initialization,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2020, pp. 696–697

    2020

  7. [7]

    Nipq: Noise proxy-based integrated pseudo-quantization,

    J. Shin, J. So, S. Park, S. Kang, S. Yoo, and E. Park, “Nipq: Noise proxy-based integrated pseudo-quantization,” inCVPR, 2023

    2023

  8. [8]

    Quan- tization meets ood: Generalizable quantization-aware training from a flatness perspective,

    J. Jiang, Y . Meng, C. Tang, H. Yu, Q. Li, Z. Wang, and W. Zhu, “Quan- tization meets ood: Generalizable quantization-aware training from a flatness perspective,” inProceedings of the 33rd ACM International Conference on Multimedia, 2025

    2025

Show all 35 references
  1. [9]

    Improving post training neural quantization: Layer-wise calibration and integer programming,

    I. Hubara, Y . Nahshan, Y . Hanani, R. Banner, and D. Soudry, “Improving post training neural quantization: Layer-wise calibration and integer programming,”arXiv preprint arXiv:2006.10518, 2020

    2006 arXiv

  2. [10]

    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 Conference on Computer Vision and Pattern Recognition, 2023, pp. 24 427–24 437. [Online]. Available: https://mlanthology.org/cvpr/2023/ liu20...

    2023

  3. [11]

    Erq: error reduction for post-training quantization of vision transformers,

    Y . Zhong, J. Hu, Y . Huang, Y . Zhang, and R. Ji, “Erq: error reduction for post-training quantization of vision transformers,” inProceedings of the 41st International Conference on Machine Learning, ser. ICML’24. JMLR.org, 2024

    2024

  4. [12]

    Quantization without tears,

    M. Fu, H. Yu, J. Shao, J. Zhou, K. Zhu, and J. Wu, “Quantization without tears,” in2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025, pp. 4462–4472

    2025

  5. [13]

    Data- free quantization through weight equalization and bias correction,

    M. Nagel, M. v. Baalen, T. Blankevoort, and M. Welling, “Data- free quantization through weight equalization and bias correction,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 1325–1334. 12

    2019

  6. [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,” in2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE Computer Society, 2020, pp. 13 166–13 175

    2020

  7. [15]

    Solving oscillation problem in post-training quantization through a theoretical perspective,

    Y . Ma, H. Li, X. Zheng, X. Xiao, R. Wang, S. Wen, X. Pan, F. Chao, and R. Ji, “Solving oscillation problem in post-training quantization through a theoretical perspective,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 7950–7959

    2023

  8. [16]

    Quantization error propagation: Revisiting layer-wise post-training quantization,

    Y . Arai and Y . Ichikawa, “Quantization error propagation: Revisiting layer-wise post-training quantization,” 2025. [Online]. Available: https://arxiv.org/abs/2504.09629

    2025

  9. [17]

    Ptq4vit: Post-training quantization for vision transformers with twin uniform quantization,

    Z. Yuan, C. Xue, Y . Chen, Q. Wu, and G. Sun, “Ptq4vit: Post-training quantization for vision transformers with twin uniform quantization,” in Proceedings of the European Conference on Computer Vision (ECCV), 2022, pp. 191–207

    2022

  10. [18]

    Fq- vit: Post-training quantization for fully quantized vision transformer,

    Y . Lin, C. Zhang, Z. Zeng, B. Ni, P. Tan, G. Xu, and C. Wang, “Fq- vit: Post-training quantization for fully quantized vision transformer,” inProceedings of the International Joint Conference on Artificial Intelligence (IJCAI), 2022, pp. 1173–1179

    2022

  11. [19]

    Repq-vit: Scale reparameterization for post-training quantization of vision transformers,

    Z. Li, B. Dong, and J. Tan, “Repq-vit: Scale reparameterization for post-training quantization of vision transformers,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 17 227–117 236

    2023

  12. [20]

    Averaging weights leads to wider optima and better generalization,

    P. Izmailov, D. Podoprikhin, T. Garipov, D. P. Vetrov, and A. G. Wilson, “Averaging weights leads to wider optima and better generalization,”CoRR, vol. abs/1803.05407, 2018. [Online]. Available: http://arxiv.org/abs/1803.05407

    2018 arXiv

  13. [21]

    Sharpness-aware minimization for efficiently improving generalization,

    P. Foret, A. Kleiner, H. Mobahi, and B. Neyshabur, “Sharpness-aware minimization for efficiently improving generalization,” inInternational Conference on Learning Representations (ICLR), 2021

    2021

  14. [22]

    Anticorre- lated noise injection for improved generalization,

    A. Orvieto, H. Kersting, F. Proske, F. Bach, and A. Lucchi, “Anticorre- lated noise injection for improved generalization,” inICML 2022-39th International Conference on Machine Learning, 2022

    2022

  15. [23]

    Bit-shrinking: A quasistatic quantization strategy for post-training quantization,

    Y . Lin, C. Xue, B. Ni, P. Tan, C. Wang, and G. Sun, “Bit-shrinking: A quasistatic quantization strategy for post-training quantization,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 20 364–20 374

    2023

  16. [24]

    Stabilizing quantization-aware training by implicit-regularization on hessian matrix,

    J. Pang and T. Cai, “Stabilizing quantization-aware training by implicit-regularization on hessian matrix,” 2025. [Online]. Available: https://arxiv.org/abs/2503.11159

    2025 arXiv

  17. [25]

    Learning multiple layers of features from tiny images,

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

    2009

  18. [26]

    Tiny imagenet visual recognition challenge,

    Y . Le and X. S. Yang, “Tiny imagenet visual recognition challenge,” Stanford CS231N Course Project, 2015

    2015

  19. [27]

    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, pp. 211–252, 2015

    2015

  20. [28]

    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

  21. [29]

    Mobilenets: Efficient convo- lutional neural networks for mobile vision applications,

    A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto, and H. Adam, “Mobilenets: Efficient convo- lutional neural networks for mobile vision applications,”arXiv preprint arXiv:1704.04861, 2017

    2017 arXiv

  22. [30]

    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

  23. [31]

    Ghostnet: More features from cheap operations,

    K. Han, Y . Wang, Q. Tian, J. Guo, C. Xu, and C. Xu, “Ghostnet: More features from cheap operations,” in2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2020

    2020

  24. [32]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” inMedical Image Computing and Computer-Assisted Intervention (MICCAI), 2015

    2015

  25. [33]

    The cityscapes dataset for semantic urban scene understanding,

    M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler, R. Be- nenson, U. Franke, S. Roth, and B. Schiele, “The cityscapes dataset for semantic urban scene understanding,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016

    2016

  26. [34]

    Pytorch: An imperative style, high-performance deep learning library,

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antigaet al., “Pytorch: An imperative style, high-performance deep learning library,”Advances in neural information processing systems, vol. 32, 2019

    2019

  27. [35]

    Mqbench: Towards reproducible and deployable model quantization benchmark,

    Y . Li, M. Shen, J. Ma, Y . Ren, M. Zhao, Q. Zhang, R. Gong, F. Yu, and J. Yan, “Mqbench: Towards reproducible and deployable model quantization benchmark,”arXiv preprint arXiv:2111.03759, 2021

    2021 arXiv

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.