Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

GDRQ: Group-based Distribution Reshaping for Quantization

T0 review · 3 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read Uniform quantization can hold near-full-precision accuracy at 2-bit weights and 4-bit activations if weights and activations are first reshaped into uniform-like distributions and each filter group learns its own quantization scale.

desk verdict A plausible engineering contribution with a clean BN-folding trick, but the ImageNet SOTA claim rests on a weak baseline, uncontrolled comparisons, and a sign error in Eq. 10 — worth refereeing, not accepting as is. read the letter →

arxiv 1908.01477 v1 pith:GACLLFKN submitted 2019-08-05 cs.CV cs.LGeess.IV

Haibao Yu , Tuopu Wen , Guangliang Cheng , Jiankai Sun , Qi Han , Jianping Shi This is my paper · ORCID
classification cs.CVcs.LGeess.IV
keywords networkquantizationuniformdistributionreshapingScale-Clipgroup-basedlow-bitneuralnetworksmodelcompression
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 sets out to establish that low-bit uniform quantization does not have to cost much accuracy, provided the network is first reshaped so its weights and activations spread out evenly across the quantization bins, and provided different filter groups get their own quantization scales. It introduces Scale-Clip, which clips weights at a multiple of their mean absolute value and keeps activations inside a dynamically updated range, and Group-based Quantization, which splits filters into groups with separate clipping values. Combined as the GDRQ framework, the paper reports a 2-bit-weight, 4-bit-activation ResNet-50 losing less than 1% top-1 accuracy on ImageNet, with similarly small losses on object detection and segmentation. The broader interest is that near-lossless low-bit models may be reachable without abandoning hardware-friendly uniform quantization.

What carries the argument

The load-bearing mechanism is the pair of Scale-Clip and Group-based Quantization. Scale-Clip is a distribution-reshaping constraint: values outside $\pm k \cdot \mathrm{mean}(|W|)$ are clipped, and when $k \approx 2$ an approximately uniform distribution emerges; for activations the threshold is updated online rather than re-estimated per batch. Group-based Quantization partitions filters into groups of size $g_s$ and solves $\alpha_l^* = \arg\min_\alpha \|G_l - Q(G_l;\alpha)\|_1 / \|W\|_1$ for each group, giving different quantization bins to different filters. Because each group's scale factor is folded into batch normalization at inference, the bit width and arithmetic stay uniform, and the model remains deployable on fixed-point hardware.

What would settle it

Run ResNet-50 on ImageNet with Scale-Clip reshaping but with one quantization scale per layer instead of per group; if 2-bit/4-bit accuracy drops more than the reported ~1 point, then the group mechanism, not distribution reshaping, is carrying the result. A second check is to measure the first-layer weight histogram after training with $k=2$: if it is not flatter than the unreshaped model's histogram, the claimed reshaping is not happening.

Watch

Extended reading notes

Core claim

The claim is that Gaussian- or Laplace-shaped weight distributions are the wrong target for uniform quantization: flatter, uniform-like distributions fit the fixed quantization bins with less quantization loss, and this shape can be induced during training without hurting the full-precision model. Scale-Clip does this by clipping each group of weights at $T_w = k \cdot \mathrm{mean}(|W|)$ with $k$ near 2, so outliers are trimmed and small values spread out, and by maintaining an activation threshold $T_a$ through the running update $T_a = T_a + \lambda(T_a - k \cdot \mathrm{mean}(|A|))$. Group-based Quantization then splits filters into groups, optimizes a separate clipping value $\alpha_l$ per group, and merges the per-group scales into the following batch-normalization layer, so inference adds no extra cost. The paper reports that this framework keeps the 2-bit/4-bit ResNet-50 within 0.9 points of its floating-point top-1 accuracy on ImageNet and keeps low-bit detection and segmentation accuracy nearly intact on PASCAL-VOC and Cityscapes.

Load-bearing premise

The load-bearing premise is that clipping a full-precision model's weights and activations to spread them evenly across the quantization bins (at roughly twice their average magnitude) leaves the model's accuracy intact; the paper's direct evidence for this premise comes from ResNet-18 on CIFAR-100, while the ImageNet results combine it with per-group quantization.

Editorial extensions

If this is right

  • ResNet-50 with 2-bit weights and 4-bit activations keeps under 1% top-1 accuracy drop on ImageNet, so near-lossless low-bit classification is within reach.
  • VGG-16-BN with 4-bit weights and 4-bit activations shows essentially no drop from its floating-point model, indicating 4-4 uniform quantization is practical on fixed-point hardware.
  • Group-based quantization raises low-bit model capacity without consuming extra bits or extra inference compute, because per-group scales merge into batch normalization.
  • The same framework transfers to detection and segmentation: a 2-bit/4-bit Faster R-CNN drops 0.7 mAP on PASCAL-VOC and a 2-bit/4-bit PSPNet drops 0.9 mIoU on Cityscapes.
  • Uniform quantization remains FPGA-friendly, unlike non-uniform schemes that require lookup tables or special arithmetic.

Reading between the lines

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

  • The paper validates reshaping alone on ResNet-18/CIFAR-100; a natural extension, not tested here, is an ImageNet ablation that turns Scale-Clip off while keeping group-based quantization, which would isolate how much of the <1% drop comes from redistribution rather than extra per-group capacity.
  • The same fixed-bin mismatch that motivates Scale-Clip appears in quantized transformers, whose activations often have large outliers; testing a running-threshold reshaping there is a direct transfer of the paper's idea to another architecture family.
  • Because per-group scales are folded into batch normalization, the method points toward co-designing quantization thresholds and normalization statistics in one end-to-end training pass, something the paper's two-stage fine-tuning procedure does not attempt.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes GDRQ, a framework for low-bit uniform quantization that combines two ingredients: Scale-Clip distribution reshaping, which clips weights at k times their mean absolute value and uses a running threshold for activations to push the layer distribution toward uniform; and Group-based Quantization, which splits convolution filters into groups with per-group clipping/scaling factors that are merged into batch normalization at inference. The central claim is a new state of the art: ResNet-50 with 2-bit weights and 4-bit activations loses less than 1% top-1 accuracy on ImageNet, with additional experiments on VGG-16, Faster R-CNN on PASCAL-VOC, and PSPNet on Cityscapes.

Significance. If confirmed, the result would be practically significant because it promises near-full-precision accuracy at very low bit widths while keeping the hardware-friendly uniform quantization format and avoiding extra inference cost by folding group scaling into batch normalization. The paper has a genuine conceptual strength: the observation that uniform-like distributions reduce per-group quantization loss is supported by a simple and correct inequality in Eq. (12), and the derivation of the 2*mean(|W|) clipping threshold from the uniform-density calculation in Eqs. (5)-(7) is clean. However, the headline state-of-the-art claim currently rests on a floating-point baseline well below the standard ResNet-50 accuracy, on comparisons copied from other papers with different training protocols, and on an ImageNet-scale transfer of assumptions that are validated only on ResNet-18/CIFAR-100.

major comments (3)
  1. [Section 3.2, Eqs. (10)-(11)] The activation threshold update as written is inconsistent with its stated objective. For the objective in Eq. (11), the gradient of (1/2)||T - k*mean(|A|)||^2 with respect to T is (T - k*mean(|A|)), so the gradient-descent update should be T <- T - lambda*(T - k*mean(|A|)). The printed update has a plus sign, yielding T <- (1+lambda)*T - lambda*k*mean(|A|), which does not converge to k*mean(|A|) and may diverge. Because the activation reshaping mechanism is load-bearing for the reported ImageNet, detection, and segmentation results, this sign error must be corrected and the experiments rerun or the text must explain a different intended update.
  2. [Section 4.3, Table 4] The central state-of-the-art claim in the abstract is relative to a floating-point ResNet-50 baseline of 74.8%, which is below the standard reported accuracy of about 76.1% for ImageNet ResNet-50. The paper attributes this to improper multi-GPU training hyperparameters, but this means the '<1% accuracy drop' is measured from an unusually low baseline, and the comparisons to SYQ, FGQ, and DoReFa-Net are taken from other papers with different training protocols and baselines, so they are not controlled comparisons. The authors should retrain or report a standard ResNet-50 baseline, state the accuracy drop relative to that baseline, and either provide a controlled reimplementation of the comparison methods or clearly restrict the claim to their own training setup.
  3. [Section 4.1, Section 4.2, Section 4.3] The distribution-reshaping assumption is validated only on ResNet-18 with CIFAR-100, yet it is silently transferred to VGG-16 and ResNet-50 on ImageNet and to detection/segmentation backbones. Table 4 reports only the combined GDRQ result; there is no ImageNet ablation that isolates the effect of Scale-Clip, and no ablation of the group-size choice at ImageNet scale. If the reshaping constraint behaves differently for large-scale models or for the activation statistics of ImageNet, the reported accuracy levels would not follow. An ablation on ImageNet (with and without Scale-Clip, and with at least the group sizes used in the final model) is needed to support the transfer.
minor comments (5)
  1. [Section 4.1, Fig. 6] In the text after Fig. 6, 'The results are consistent with the results shown in Fig. 6(b)' should probably refer to Fig. 6(a), since the discussion is about quantized-loss before mentioning accuracy.
  2. [Table 2] The entries '0.1 0.1 0.1 0.1' and '-1.6 -1.4 -1.4 -1.7' are labeled as Top-1 Accuracy but appear to be accuracy drops relative to the float model; the table header and caption should be clarified.
  3. [Eq. (12)] The per-group quantization loss divides by ||W||_1, the norm of the full layer, while the numerator is the group loss; since the denominator is constant across groups, this is harmless, but the notation should be clarified to avoid implying a group-specific denominator.
  4. [Conclusion] The conclusion mentions experiments on 'COCO' and 'COCO' in the text, but the reported detection experiments are on PASCAL-VOC; the text should be corrected to match the experimental section.
  5. [References] Several references are incomplete or malformed, for example Ref. [13] 'C. J., W. Z., S. Venkataramani, C. P. I., S. V., and G. K.'; these should be reconstructed and checked.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the Scale-Clip and group-quantization derivations are self-contained, and the SOTA claim is benchmarked against external methods.

full rationale

The paper's derivation chain is not circular. Scale-Clip is motivated by an external, independently testable observation (Han et al.) that uniform distributions quantize better and by an explicit calculation (Eqs. 5-7) showing max(|W|) is approximately 2*mean(|W|) for a uniform distribution; the clipping rule T_w = k*mean(|W|) is an ansatz, not a fit to the reported accuracy. The group-quantization inequality in Section 3.3 is a genuine monotonicity property: splitting W into groups and choosing alpha_i per group can only lower the sum of per-group quantization losses relative to one common alpha, because each per-group minimizer is no worse than the common-alpha loss for that group. The ImageNet SOTA claim is tested against external methods (SYQ, FGQ, DoReFa), not derived from those methods. The validation in Section 4.1 independently varies k and measures both the quantization loss (Eq. 3) and the fine-tuned Top-1 accuracy, so the performance comparison is not forced by construction. The only self-citation is the PSPNet architecture reference [32], which is used as an external backbone and is not load-bearing for the quantization derivation. The one correctness concern, that the activation threshold update in Eq. 10 as written has a sign inconsistent with the stated minimization in Eq. 11, would affect reproducibility but does not constitute circularity, because the target T = k*mean(|A|) is not the paper's accuracy claim. Overall, the central claims are supported by independent experiments and external comparisons rather than by self-referential reasoning.

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

The method introduces no new physical entities or forces. The load-bearing assumptions are about the relationship between weight/activation distribution shape and quantizer performance, plus the validity of using reconstruction loss as a proxy for task accuracy. The free parameters are the clipping factor, group size, activation threshold step size, and the per-group fitted thresholds.

free parameters (4)
  • Scale-Clip factor k = 2 (default), candidates 2, 2.5, 3, 4
    Chosen by hand from the uniform-distribution identity T=2 mean(|W|) and validated in Section 4.1. It is a hyperparameter, not independently derived for activations.
  • Group size gs = 1 for CIFAR classification experiments; values 1, 4, 16, -1 explored; not stated for ImageNet experiments
    Selected per experiment in Section 4.2 to balance capacity and stability. The paper explicitly says a 'proper group size' must be chosen, so the method depends on this selection.
  • Activation update step size lambda = not specified
    Used in Eq. 10 to update the activation clipping threshold T_a. No value or schedule is given, which is needed to reproduce the method.
  • Per-group clipping thresholds alpha_l = optimized per group via Eq. 12
    Each filter group gets a separate quantization scale fitted to minimize reconstruction loss. These values are data-dependent and are central to the group-based method.
assumptions (4)
  • domain assumption Uniform-like distributions of weights and activations are more amenable to low-bit uniform quantization than Gaussian or Laplace-like distributions.
    Used to motivate Scale-Clip in Section 3.2 and Fig. 1; supported only by a toy example and reference [11].
  • domain assumption Forcing weights to a uniform-like distribution via clipping at k times the mean absolute value does not materially hurt full-precision model accuracy.
    Assumed in Section 3.2; tested only for ResNet-18 on CIFAR-100 in Section 4.1, then applied to ImageNet, detection, and segmentation.
  • domain assumption Quantization loss ||W - Q(W;alpha)|| / ||W|| is a reliable proxy for final task accuracy, and minimizing it per group improves the low-bit model.
    Used in Eq. 4 and Eq. 12 to select all alpha; no theoretical connection to classification loss is shown.
  • domain assumption The gradient through the quantizer can be obtained via straight-through estimation and the quantized network can be fine-tuned with standard SGD.
    Presumed throughout the experiments; straight-through estimation is not described in the method section, only referenced indirectly via [35].

how reviews work

0 comments
Cite this review

Pith. "Pith review of GDRQ: Group-based Distribution Reshaping for Quantization." pith.science (2026). https://pith.science/paper/GACLLFKN

@misc{pith2026190801477,
  author       = {Pith},
  title        = {Pith review of: GDRQ: Group-based Distribution Reshaping for Quantization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GACLLFKN}},
  note         = {Machine review of arXiv:1908.01477}
}
read the original abstract

Low-bit quantization is challenging to maintain high performance with limited model capacity (e.g., 4-bit for both weights and activations). Naturally, the distribution of both weights and activations in deep neural network are Gaussian-like. Nevertheless, due to the limited bitwidth of low-bit model, uniform-like distributed weights and activations have been proved to be more friendly to quantization while preserving accuracy~\cite{Han2015Learning}. Motivated by this, we propose Scale-Clip, a Distribution Reshaping technique that can reshape weights or activations into a uniform-like distribution in a dynamic manner. Furthermore, to increase the model capability for a low-bit model, a novel Group-based Quantization algorithm is proposed to split the filters into several groups. Different groups can learn different quantization parameters, which can be elegantly merged in to batch normalization layer without extra computational cost in the inference stage. Finally, we integrate Scale-Clip technique with Group-based Quantization algorithm and propose the Group-based Distribution Reshaping Quantization (GDQR) framework to further improve the quantization performance. Experiments on various networks (e.g. VGGNet and ResNet) and vision tasks (e.g. classification, detection and segmentation) demonstrate that our framework achieves good performance.

Figures

Figures reproduced from arXiv: 1908.01477 by the authors.

Figure 1
Figure 1. We respectively quantize the Ws which obey: (a) Laplace-like distribution, (b) Gaussian-like distribution and (c) Uniform-like distribution into Q(W) as Eq. 12 with nw = 4 and || · ||p = || · ||1, where red lines means the quantization bins. We calculate quantized-loss as Eq. 3. Comparatively, uniform distribution better fits the uniform quantization. restricts them to be deployment on resource-limited de￾vices (e.g… view at source ↗
Figure 2
Figure 2. Overview of our quantization framework. (a) illustrates the main flow of proposed group-based quantization. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Shape the Gaussian-like distribution into [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Optimal α for first convolutional layer’s weights W of trained ResNet-18. Blue bar represents the optimal α ∗ for W while orange bars are correspond to each group filters. our work, we just took advantage of the clipped method as part of our optimization, and compared …
Figure 5
Figure 5. Figure 5: The blue bars show the weight distribution of the [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: (a) Quantized-loss of the first convolutional [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Loss and Accuracy with different group size [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Evolution Strategy-Based Calibration for Low-Bit Quantization of Speech Models

    cs.SD 2026-03 conditional novelty 5.0 of 10

    Using CMA-ES to jointly optimize activation quantization scales keeps speech-model accuracy near full precision under full INT8 and INT4 quantization.

Reference graph

Works this paper leans on

38 extracted references · 23 canonical work pages · cited by 1 Pith paper

  1. [1]

    G. I. e. a. Abadi M, Chu A. Deep learning with differential privacy. pages 308–318, 2016. 4

    work page 2016

  2. [2]

    Z. Cai, X. He, J. Sun, and N. Vasconcelos. Deep learn- ing with low precision by half-wave gaussian quantization. arXiv preprint arXiv:1702.00953, 2017. 2, 7

    arXiv 2017

  3. [3]

    Cheng, P.-s

    J. Cheng, P.-s. Wang, G. Li, Q.-h. Hu, and H.-q. Lu. Re- cent advances in efficient computation of deep convolu- tional neural networks. Frontiers of Information Technology & Electronic Engineering, 19(1):64–77, 2018. 2

    work page 2018

  4. [4]

    Courbariaux, I

    M. Courbariaux, I. Hubara, D. Soudry, R. El-Yaniv, and Y . Bengio. Binarized neural networks: Training deep neu- ral networks with weights and activations constrained to+ 1 or-1. arXiv preprint arXiv:1602.02830, 2016. 1, 2

    arXiv 2016

  5. [5]

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei- Fei. Imagenet: A large-scale hierarchical image database. In Computer Vision and Pattern Recognition, 2009. CVPR

    2009

  6. [6]

    E. L. Denton, W. Zaremba, J. Bruna, Y . LeCun, and R. Fer- gus. Exploiting linear structure within convolutional net- works for efficient evaluation. In Advances in neural infor- mation processing systems, pages 1269–1277, 2014. 1

    work page 2014

  7. [7]

    Faraone, N

    J. Faraone, N. Fraser, M. Blott, and P. H. Leong. Syq: Learning symmetric quantization for efficient deep neural networks. In Proceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition , pages 4300–4309,

    work page

  8. [8]

    Girshick

    R. Girshick. Fast r-cnn. In Proceedings of the IEEE inter- national conference on computer vision , pages 1440–1448,

Show all 38 references
  1. [9]

    P. Gysel. Ristretto: Hardware-oriented approxima- tion of convolutional neural networks. arXiv preprint arXiv:1605.06402, 2016. 1, 2

    2016 arXiv

  2. [10]

    S. Han, H. Mao, and W. J. Dally. Deep compres- sion: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015. 1, 2, 4, 5

    2015 arXiv

  3. [11]

    S. Han, J. Pool, J. Tran, and W. J. Dally. Learning both weights and connections for efficient neural networks. pages 1135–1143, 2015. 1

    2015

  4. [12]

    K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learn- ing for image recognition. In Proceedings of the IEEE con- ference on computer vision and pattern recognition , pages 770–778, 2016. 1

    2016

  5. [13]

    C. J., W. Z., S. Venkataramani, C. P. I., S. V ., and G. K. Pact: Parameterized clipping activation for quantized neural networks. arXiv preprint arXiv:1805.06085, 2018. 3, 4, 5

    2018 arXiv

  6. [14]

    H. L., Z. R., and J. T. Kwok. Analysis of quantized models

  7. [15]

    J. Long, E. Shelhamer, and T. Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 3431–3440, 2015. 1

    2015

  8. [16]

    Mellempudi, A

    N. Mellempudi, A. Kundu, D. Mudigere, D. Das, B. Kaul, and P. Dubey. Ternary neural networks with fine-grained quantization. arXiv preprint arXiv:1705.01462, 2017. 2

    2017 arXiv

  9. [17]

    Mishra and D

    A. Mishra and D. Marr. Apprentice: Using knowledge dis- tillation techniques to improve low-precision network accu- racy. arXiv preprint arXiv:1711.05852, 2017. 1

    2017 arXiv

  10. [18]

    Miyashita, E

    D. Miyashita, E. H. Lee, and B. Murmann. Convolu- tional neural networks using logarithmic data representa- tion. arXiv preprint arXiv:1603.01025, 2016. 2

    2016 arXiv

  11. [19]

    L. Q. V . e. a. Neelakantan A, Vilnis L. Adding gradient noise improves learning for very deep networks. arXiv preprint arXiv:1511.06807, 2015. 4

    2015 arXiv

  12. [20]

    Y . P., L. J., S. Zhang, O. S., Q. Y ., and X. J. Understanding straight-through estimator in training activation quantized neural nets. 2019. 3

    2019

  13. [21]

    E. Park, J. Ahn, and S. Yoo. Weighted-entropy-based quan- tization for deep neural networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2, 8

    2017

  14. [22]

    Rastegari, V

    M. Rastegari, V . Ordonez, J. Redmon, and A. Farhadi. Xnor-net: Imagenet classification using binary convolu- tional neural networks. In European Conference on Com- puter Vision, pages 525–542. Springer, 2016. 2

    2016

  15. [23]

    Redmon, S

    J. Redmon, S. Divvala, R. Girshick, and A. Farhadi. You only look once: Unified, real-time object detection. In Pro- ceedings of the IEEE conference on computer vision and pattern recognition, pages 779–788, 2016. 1

    2016

  16. [24]

    S. Ren, K. He, R. Girshick, and J. Sun. Faster r-cnn: To- wards real-time object detection with region proposal net- works. In Advances in neural information processing sys- tems, pages 91–99, 2015. 1

    2015

  17. [25]

    Sze, Y .-H

    V . Sze, Y .-H. Chen, T.-J. Yang, and J. S. Emer. Efficient processing of deep neural networks: A tutorial and survey. Proceedings of the IEEE, 105(12):2295–2329, 2017. 2

    2017

  18. [26]

    K. Wang, Z. Liu, Y . Lin, J. Lin, and S. Han. Haq: Hardware-aware automated quantization. arXiv preprint arXiv:1811.08886, 2018. 3, 4

    2018 arXiv

  19. [27]

    Wang and J

    P. Wang and J. Cheng. Fixed-point factorized networks. In Computer Vision and Pattern Recognition (CVPR), 2017 IEEE Conference on, pages 3966–3974. IEEE, 2017. 2

    2017

  20. [28]

    P. Wang, Q. Hu, Y . Zhang, C. Zhang, Y . Liu, J. Cheng, et al. Two-step quantization for low-bit neural networks. In Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4376–4384, 2018. 1, 2

    2018

  21. [29]

    B. Wu, Y . Wang, P. Zhang, Y . Tian, P. Vajda, and K. Keutzer. Mixed precision quantization of convnets via differentiable neural architecture search. arXiv preprint arXiv:1812.00090, 2018. 3

    2018 arXiv

  22. [30]

    J. Wu, C. Leng, Y . Wang, Q. Hu, and J. Cheng. Quantized convolutional neural networks for mobile devices. In Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4820–4828, 2016. 2

    2016

  23. [31]

    P. Yin, S. Zhang, Y . Qi, and J. Xin. Quantization and train- ing of low bit-width convolutional neural networks for ob- ject detection. arXiv preprint arXiv:1612.06052, 2016. 8

    2016 arXiv

  24. [32]

    H. Zhao, J. Shi, X. Qi, X. Wang, and J. Jia. Pyramid scene parsing network. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 2881–2890, 2017. 1

    2017

  25. [33]

    A. Zhou, A. Yao, Y . Guo, L. Xu, and Y . Chen. Incremen- tal network quantization: Towards lossless cnns with low- precision weights. arXiv preprint arXiv:1702.03044, 2017. 1, 2

    2017 arXiv

  26. [34]

    A. Zhou, A. Yao, K. Wang, and Y . Chen. Explicit loss- error-aware quantization for low-bit deep neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9426–9435, 2018. 2

    2018

  27. [35]

    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. 1, 2

    2016 arXiv

  28. [36]

    Zhou, Y .-Z

    S.-C. Zhou, Y .-Z. Wang, H. Wen, Q.-Y . He, and Y .-H. Zou. Balanced quantization: An effective and efficient approach to quantized neural networks. Journal of Computer Science and Technology, 32(4):667–682, 2017. 2

    2017

  29. [37]

    C. Zhu, S. Han, H. Mao, and W. J. Dally. Trained ternary quantization. 2016. 1 9

    2016

  30. [2009]

    Ieee, 2009

    IEEE Conference on, pages 248–255. Ieee, 2009. 1 8

    2009

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.