Pith. sign in

REVIEW 4 major objections 6 minor 52 references

Combining geometric-median pruning with APoT quantization compresses CIFAR-10 CNNs by roughly 15x while keeping accuracy within about two points of the full-precision baseline.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-05 10:13 UTC pith:BMMAXM2K

load-bearing objection Honest combination of two known techniques with plausible CIFAR-10 numbers, but the paper never explains how 30% per-layer pruning handles ResNet skip connections, so the headline 15x compression is not reproducible as written. the 4 major comments →

arxiv 2509.04244 v1 pith:BMMAXM2K submitted 2025-09-04 cs.NE

Integrating Pruning with Quantization for Efficient Deep Neural Networks Compression

Sara Makenali , Babak Rokh , Ali Azarpeyvand This is my paper
classification cs.NE
keywords model compressionfilter pruninggeometric medianAPoT quantizationpower-of-two quantizationquantization-aware trainingCIFAR-10deep neural networks
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper sets out to show that two compression techniques usually applied separately—similarity-based filter pruning and low-bit quantization—work better when deliberately integrated. It proposes two integration recipes: SPQ, which prunes and quantizes together during every training epoch, and PPQ, which first prunes incrementally and only then applies 4-bit quantization-aware training. On CIFAR-10 ResNet-20/32/56/110 and VGG-16, the authors report roughly a 15x cut in model size and more than a 100x cut in bit-operations, with accuracy losses around 0.1 to 2.2 percentage points, and with PPQ slightly exceeding the full-precision baseline on ResNet-110. If these results hold, the recipes are practical ways to deploy accurate image classifiers on devices with limited memory and compute.

Core claim

The central claim is that combining GM-based filter pruning with APoT quantization yields effective compression without the accuracy collapse usually feared from stacking two lossy operations. SPQ quantizes weights and activations to 4-bit APoT in every epoch and zero-masks the 30% of filters closest to each layer's geometric median at the end of each epoch, while continuing to update the pruned filters. PPQ trains a full-precision network, prunes in two incremental stages (30% total) using the same GM criterion, then performs quantization-aware training for about 50 epochs. The reported outcome is that PPQ is the more accurate pipeline on all four ResNets—best among compared methods on ResN

What carries the argument

Geometric-median filter pruning: in each convolutional layer, the GM is the point minimizing total Euclidean distance to all filters; the filters nearest it are judged redundant and zero-masked, because their represented information is assumed to overlap with the remaining filters. Additive Powers-of-Two (APoT) quantization: quantization levels are sums of n powers of two, matching the bell-shaped weight and activation distribution; at 4-bit precision, MAC multiplications can be replaced by bit-shift operations. First and last layers are kept at 8-bit. The two integration schemes—SPQ's simultaneous application each epoch and PPQ's staged pruning followed by quantization-aware training—carry

Load-bearing premise

The load-bearing premise is that a filter's distance from the geometric median of full-precision weights still identifies redundant filters after weights and activations are quantized to 4-bit, even though the forward pass never sees the full-precision weights.

What would settle it

Run both pipelines on CIFAR-10 with the same 30% pruning rate and the same 4-bit APoT settings, but replace the GM importance score with random filter selection or with smallest-L2-norm filters. If random pruning reproduces the reported accuracy and compression trade-offs, the geometric-median criterion is not load-bearing; if GM-selected pruning consistently beats the controls across ResNet and VGG, the criterion is doing the claimed work.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • PPQ on ResNet-110 reaches 94.56% accuracy, 0.06 points above the full-precision baseline, while cutting model size by x15.94 and BOPs by x120.95.
  • On all tested ResNets, PPQ stays within 0.77 points of baseline on ResNet-20 and often beats pruning-only or quantization-only SOTA methods, making sequential integration the safer default for deep residual networks.
  • SPQ achieves the best accuracy among all compared methods on VGG-16 (93.81%, 0.24-point drop) at a lower training cost than PPQ, making simultaneous integration attractive for wide, overparameterized networks.
  • Because both weights and activations are APoT-quantized, inference can replace floating-point multiplications with bit shifts, so the savings go beyond model size.
  • Structured filter pruning also removes the corresponding input channels in the next layer, keeping the compressed model hardware-friendly.

Where Pith is reading between the lines

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

  • The paper pairs one redundancy-based pruning score with one low-bit quantizer, but the same integration recipe could be tested with other similarity-based pruning criteria and other non-uniform quantizers; the paper only demonstrates the GM-plus-APoT combination.
  • The SPQ-versus-PPQ accuracy pattern hints at a rule the authors do not state: sequential compression protects representation in deep residual networks with narrow bottlenecks, while simultaneous compression exploits redundancy in heavily overparameterized networks like VGG-16.
  • BOPs counts assume bit-width-proportional arithmetic cost; actual latency and energy gains on a given chip also depend on memory traffic and whether shift-add units are implemented, so the reported x115–126 reductions are not automatically the same as wall-clock speedups.
  • Uniform 30% pruning and uniform 4-bit precision were fixed across all layers; layer-wise per-sensitivity pruning rates and bit-widths, mentioned only as future work, are the natural next lever and could push compression beyond x15 at the same accuracy.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes two pipelines that combine GM-based structured filter pruning with APoT low-bit quantization: SPQ, which applies pruning and quantization simultaneously during training, and PPQ, which first prunes and then performs quantization-aware training. Experiments on CIFAR-10 with ResNet-20/32/56/110 and VGG-16 report roughly ×15.8–×15.95 model-size reduction and ×115–×126 BOPs reduction with small accuracy drops, and the authors claim PPQ achieves the best accuracy on ResNet-32 and ResNet-110 while SPQ is best on VGG-16. The paper is an empirical integration of existing techniques; it provides no new theoretical derivation.

Significance. If the reported results are reproducible, the two pipelines would be practically useful: combining 30% structured pruning with 4-bit APoT quantization yields roughly a further 2× compression over quantization alone, with reported accuracy drops mostly between 0.12% and 2.21%. The manuscript is strengthened by reporting comparisons against several SOTA methods on public benchmarks and by defining the compression metrics explicitly. However, the contribution is primarily incremental—GM pruning and APoT quantization are both published methods—and several load-bearing experimental details are missing. The paper also ships no code or machine-checked artifacts, so the credibility of the headline numbers rests entirely on the textual description, which is currently incomplete.

major comments (4)
  1. [Section III-A, Algorithms 1–2, Eq. (5), Table II] The pruning procedure is not specified for residual networks. Both algorithms select filters per convolutional layer and zero-mask them, but neither describes how pruning interacts with ResNet skip connections. If 30% of filters are removed from both conv layers in a basic block, the block's output channel count no longer matches the shortcut; if only the first conv is pruned, the stated 'all convolutional layers, 30%' policy and the computed compression ratios change. If the filters are only zero-masked and the channels are kept, then Eq. (5) counts those zeroed weights, so the reported ×15.78–×15.94 model-size reductions (and corresponding BOPs reductions) do not follow. The authors must state the exact channel-removal rule for ResNet blocks, including downsampling shortcuts, and report the compression metrics for the actually deployed architecture.
  2. [Section IV-A, Algorithms 1–2, Eq. (4)] Key hyperparameters are never reported. The number of training epochs n, the number of QAT epochs m, the number of pruning stages s, the stage-wise pruning schedule {p_i}, and the APoT parameters k and α are all essential for reproducibility, but Section IV only gives learning-rate schedules, batch size, and the global 30% pruning ratio. For example, Algorithm 2's pruning condition depends on n and s, and the quantization levels in Eq. (4) are determined by k and α. Please provide a complete hyperparameter table for every architecture and both methods.
  3. [Tables II–III, Fig. 8] Several headline claims rest on accuracy differences that are within plausible run-to-run variation of a single CIFAR-10 training run. For example, PPQ on ResNet-110 is reported as 94.56% versus a 94.50% baseline, and PPQ versus SPQ on VGG-16 differs by 0.35 percentage points. No error bars, number of seeds, or statistical significance tests are reported. Since the central claim is that PPQ/SPQ match or exceed SOTA accuracy while achieving much larger compression, at least 3–5 independent runs with mean ± std (and, ideally, paired significance tests) are needed to support the comparison.
  4. [Section III-C, Fig. 4] The pruning criterion is computed on full-precision weights even though the forward pass in SPQ uses 4-bit APoT-quantized weights. The paper explicitly says 'GM is computed based on full-precision weights' but does not justify why the geometric center of full-precision filters is the right redundancy signal for a low-bit quantized model. If filter redundancy changes after quantization, the pruning decisions—and therefore the reported accuracy/compression trade-off—may not generalize. Please provide an ablation comparing GM computed on full-precision weights against GM computed on quantized weights, or otherwise justify the mismatch.
minor comments (6)
  1. [Author block / Abstract] The first author's name is spelled 'Sara Makenali' in the abstract and 'Sara Mekenali' in the author block; please unify.
  2. [Section IV, first paragraph] Typo: 'ReNet-32' should be 'ResNet-32'.
  3. [Fig. 1] Typo in the figure: 'Quatization' should be 'Quantization'.
  4. [Algorithm 2, line 23] The loop header 'For epoch = 1 to m ⌈|D|/BS⌉' is malformed; presumably the quantization phase should iterate over epochs and, inside each epoch, over mini-batches.
  5. [Fig. 8] The x-axis labels are garbled/unreadable in the submitted PDF; please replace with clean text.
  6. [Section III-B, Eq. (4)] The APoT notation is introduced but the specific values of k, α, and n used in the 4-bit experiments are not stated; please add them in Section IV.

Circularity Check

0 steps flagged

No significant circularity: the paper empirically combines published GM pruning and APoT quantization; the only self-citation is minor and non-load-bearing.

full rationale

The paper makes no formal derivation that could reduce to its inputs. Its central claims are empirical: two training schedules (SPQ and PPQ) that combine FPGM's geometric-median filter pruning [20] with APoT quantization [4], evaluated on CIFAR-10 against external SOTA baselines. The reported compression ratios in Tables II and III follow arithmetically from Eqs. (5)-(6) with the stated 30% pruning rate and 4-bit precision; they are metric computations, not predictions fitted to the data. Accuracy numbers are measured, not derived from the methods' assumptions. The only self-citation is [16], the authors' own quantization survey, used for background statements about quantization and PoT/FPGA suitability; it does not justify the central claim, exclude alternatives, or supply a uniqueness theorem. GM pruning and APoT quantization are imported from external publications [20], [4], so no ansatz is smuggled in via self-citation. The possible missing handling of ResNet skip connections in the pruning description is a reproducibility/correctness concern, not an input-output circularity. Thus no circular step is identified; the score reflects only the presence of the minor non-load-bearing self-citation.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The paper introduces no new mathematical object or entity. It relies on two published techniques (FPGM and APoT) and standard training hyperparameters. The most significant unstated inputs are the specific hyperparameter values (epoch counts, APoT settings) that the experiments depend on but that are not reported.

free parameters (6)
  • global pruning ratio p = 30%
    Set for all convolutional layers of every network; no sensitivity analysis or per-layer adaptation is reported (Section IV).
  • weight/activation bit-width = 4-bit for all layers except first and last at 8-bit
    Chosen by hand; no ablation of 3-bit or 5-bit alternatives is provided (Section IV).
  • APoT hyperparameters (k, alpha) = not reported
    APoT from [4] requires base bit-width k and clipping threshold alpha; the paper specifies b=4 and n=b/k but never reports k or alpha (Section III-B).
  • training epochs (n, m) and pruning stages (s) = not reported
    Algorithms 1 and 2 depend on n, m, and s; only convergence plots suggest roughly 200 epochs for ResNet and 150 for VGG, with QAT around 50 epochs (Figure 7).
  • learning rate schedules = initial 0.1, decay by 0.9 every 3 epochs if no improvement; QAT uses 0.01
    Manual schedule choices that affect final accuracy and convergence (Section IV-A).
  • optimizer hyperparameters = batch size 128, momentum 0.9, weight decay 5e-4
    Standard values chosen by hand and applied uniformly; not analyzed for sensitivity.
axioms (4)
  • domain assumption Filters closest to the geometric median of a layer's filters are redundant and can be pruned without significant accuracy loss (from FPGM [20]).
    Adopted as the pruning criterion in both SPQ and PPQ (Section III-A); the paper does not re-derive this property and relies on its validity under 4-bit quantization.
  • domain assumption APoT quantization with additive powers of two can represent DNN weights and activations at 4-bit with acceptable error (from [4]).
    Used in both methods (Section III-B); the paper does not verify quantization error or distribution fit on its own models beyond final accuracy.
  • standard math The BOPs metric (Eq. 6) is a valid proxy for computational cost when comparing compression methods.
    Used in compression comparison (Section IV-B); standard in the field but an assumption about hardware.
  • domain assumption CIFAR-10 accuracy is representative of DNN compression performance for deployment on resource-constrained devices.
    All experiments are on CIFAR-10; generalization to ImageNet or other tasks is not demonstrated.

pith-pipeline@v1.4.0-alltime-deepseek-medium · 17712 in / 13823 out tokens · 108005 ms · 2026-08-05T10:13:09.163823+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Integrating Pruning with Quantization for Efficient Deep Neural Networks Compression." pith.science (2026). https://pith.science/paper/BMMAXM2K

@misc{pith2026250904244,
  author       = {Pith},
  title        = {Pith review of: Integrating Pruning with Quantization for Efficient Deep Neural Networks Compression},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BMMAXM2K}},
  note         = {Machine review of arXiv:2509.04244}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Deep Neural Networks (DNNs) have achieved significant advances in a wide range of applications. However, their deployment on resource-constrained devices remains a challenge due to the large number of layers and parameters, which result in considerable computational and memory demands. To address this issue, pruning and quantization are two widely used compression techniques, commonly applied individually in most studies to reduce model size and enhance processing speed. Nevertheless, combining these two techniques can yield even greater compression benefits. Effectively integrating pruning and quantization to harness their complementary advantages poses a challenging task, primarily due to their potential impact on model accuracy and the complexity of jointly optimizing both processes. In this paper, we propose two approaches that integrate similarity-based filter pruning with Adaptive Power-of-Two (APoT) quantization to achieve higher compression efficiency while preserving model accuracy. In the first approach, pruning and quantization are applied simultaneously during training. In the second approach, pruning is performed first to remove less important parameters, followed by quantization of the pruned model using low-bit representations. Experimental results demonstrate that our proposed approaches achieve effective model compression with minimal accuracy degradation, making them well-suited for deployment on devices with limited computational resources.

Figures

Figures reproduced from arXiv: 2509.04244 by Ali Azarpeyvand, Babak Rokh, Sara Makenali.

Figure 1
Figure 1. Figure 1: Simultaneous pruning and quantization in SPQ, and [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Weight distributions in trained models. [31]. For instance, [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: An overview of the SPQ method [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The processing pipeline of weights in a training iter [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: presents an overview of this method, in which incre￾mental pruning is performed before quantization. Instead of pruning all filters at once, the pruning rate gradually increases. Initially, the full-precision network is trained for n epochs, after which p% of the filters are pruned, where p < pmax, and Backward and Update Unpruned Weights Quantized Unpruned Weights Quantized Activations Convolution Forward… view at source ↗
Figure 6
Figure 6. Figure 6: Illustration of the weight pruning and quantization [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Training accuracy convergence curves for SPQ, PPQ, a [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of accuracy drop relative to the uncom [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Comparison of model sizes and BOPs among the proposed [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

52 extracted references · 42 canonical work pages · 6 internal anchors

  1. [1]

    M. A. Wani, S. Ali, M. A. Sofi, and B. Sultan, Advances in Deep Learning, V olume 2. Springer, 2025, vol. 12

    work page 2025

  2. [2]

    Adam: Adaptive ap proximate multiplier for fault tolerance in dnn accelerators,

    M. Taheri, N. Cherezova, S. Nazari, A. Azarpeyvand, T. Gh asempouri, M. Daneshtalab, J. Raik, and M. Jenihhin, “Adam: Adaptive ap proximate multiplier for fault tolerance in dnn accelerators,” IEEE Transactions on Device and Materials Reliability , vol. 25, no. 1, pp. 66–75, 2025

    work page 2025

  3. [3]

    T ransaxx: Efficient transformers with approximate computing,

    D. Danopoulos, G. Zervakis, D. Soudris, and J. Henkel, “T ransaxx: Efficient transformers with approximate computing,” IEEE Transactions on Circuits and Systems for Artificial Intelligence , 2025

    work page 2025

  4. [4]

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

    Y . Li, X. Dong, and W. Wang, “Additive powers-of-two quan tization: An efficient non-uniform discretization for neural networks, ” arXiv preprint arXiv:1909.13144, 2019

    Pith/arXiv arXiv 1909

  5. [5]

    NUPES : Non-Uniform Post-Training Quantization via Power Exponent Search

    E. Yvinec, A. Dapogny, and K. Bailly, “Nupes : Non-unifor m post-training quantization via power exponent search,” 20 23. [Online]. Available: https://arxiv.org/abs/2308.05600

    work page internal anchor Pith review Pith/arXiv arXiv

  6. [6]

    Communication-Efficient Federated Learning via Clipped Uniform Quantization

    Z. Bozorgasl and H. Chen, “Clipped uniform quantizers fo r communication-efficient federated learning,” 2024. [Onli ne]. Available: https://arxiv.org/abs/2405.13365

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  7. [7]

    Quantizat ion without tears,

    M. Fu, H. Y u, J. Shao, J. Zhou, K. Zhu, and J. Wu, “Quantizat ion without tears,” in Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 4462–4472

    work page 2025

  8. [8]

    Snip: Single-shot ne twork pruning based on connection sensitivity,

    N. Lee, T. Ajanthan, and P . H. Torr, “Snip: Single-shot ne twork pruning based on connection sensitivity,” arXiv preprint arXiv:1810.02340, 2018

    Pith/arXiv arXiv 2018

  9. [9]

    Efficient CNNs via Passive Filter Pruning

    A. Singh and M. D. Plumbley, “Efficient cnns via passive fil ter pruning,” 2023. [Online]. Available: https://arxiv.org/ abs/2304.02319

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  10. [10]

    Consecutive layer collaborati ve filter similarity for differentiable neural network pruning,

    X. Zu, Y . Li, and B. Yin, “Consecutive layer collaborati ve filter similarity for differentiable neural network pruning,” Neurocomputing, vol. 533, pp. 35–45, 2023

    work page 2023

  11. [11]

    Pruning convolution neural n etworks using filter clustering based on normalized cross-correlat ion similarity,

    N. A. Khan and A. S. Rafat, “Pruning convolution neural n etworks using filter clustering based on normalized cross-correlat ion similarity,” Journal of Information and Telecommunication , vol. 9, no. 2, pp. 190– 208, 2025

    work page 2025

  12. [12]

    Losparse: Structured compression of large language model s based on low-rank and sparse approximation,

    Y . Li, Y . Y u, Q. Zhang, C. Liang, P . He, W. Chen, and T. Zhao , “Losparse: Structured compression of large language model s based on low-rank and sparse approximation,” in International Conference on Machine Learning . PMLR, 2023, pp. 20 336–20 350

    work page 2023

  13. [13]

    Low-Rank Matrix Approximation for Neural Network Compression

    K. Cherukuri and A. Lala, “Low-rank matrix approximati on for neural network compression,” arXiv preprint arXiv:2504.20078 , 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  14. [14]

    Uncertai nty-based knowledge distillation for bayesian deep neural network co mpression,

    M. Hemmatian, A. Shahzadi, and S. Mozaffari, “Uncertai nty-based knowledge distillation for bayesian deep neural network co mpression,” International Journal of Approximate Reasoning , vol. 175, p. 109301, 2024. 12

    work page 2024

  15. [15]

    Counterclockwise block-by-block knowledge distillatio n for neural network compression,

    X. Lan, Y . Zeng, X. Wei, T. Zhang, Y . Wang, C. Huang, and W. He, “Counterclockwise block-by-block knowledge distillatio n for neural network compression,” Scientific Reports, vol. 15, no. 1, p. 11369, 2025

    work page 2025

  16. [16]

    A compre hensive survey on model quantization for deep neural networks in ima ge classification,

    B. Rokh, A. Azarpeyvand, and A. Khanteymoori, “A compre hensive survey on model quantization for deep neural networks in ima ge classification,” ACM Trans. Intell. Syst. Technol. , vol. 14, no. 6, pp. 1–50, Nov. 2023

    work page 2023

  17. [17]

    Hfpq: deep neural network com pression by hardware-friendly pruning-quantization,

    Y . Fan, W. Pang, and S. Lu, “Hfpq: deep neural network com pression by hardware-friendly pruning-quantization,” Applied Intelligence, pp. 1–13, 2021

    work page 2021

  18. [18]

    Hardware-aware dnn compression via diverse prun- ing and mixed-precision quantization,

    K. Balaskas, A. Karatzas, C. Sad, K. Siozios, I. Anagnos topoulos, G. Zervakis et al. , “Hardware-aware dnn compression via diverse prun- ing and mixed-precision quantization,” IEEE Transactions on Emerging Topics in Computing , 2024

    work page 2024

  19. [19]

    Optimized convolutional ne ural network at the iot edge for image detection using pruning and quantiz ation,

    S. Naveen and M. R. Kounte, “Optimized convolutional ne ural network at the iot edge for image detection using pruning and quantiz ation,” Multimedia Tools and Applications, vol. 84, no. 9, pp. 5435–5455, 2025

    work page 2025

  20. [20]

    Filter pruning via geometric median for deep convolutional neural networks ac celeration,

    Y . He, P . Liu, Z. Wang, Z. Hu, and Y . Y ang, “Filter pruning via geometric median for deep convolutional neural networks ac celeration,” in Proceedings of the IEEE/CVF conference on computer vision a nd pattern recognition, 2019, pp. 4340–4349

    work page 2019

  21. [21]

    Differentiable joi nt pruning and quantization for hardware efficiency,

    Y . Wang, Y . Lu, and T. Blankevoort, “Differentiable joi nt pruning and quantization for hardware efficiency,” in European Conference on Computer Vision. Springer, 2020, pp. 259–277

    work page 2020

  22. [22]

    Learning both wei ghts and con- nections for efficient neural network,

    S. Han, J. Pool, J. Tran, and W. Dally, “Learning both wei ghts and con- nections for efficient neural network,” Advances in neural information processing systems, vol. 28, 2015

    work page 2015

  23. [23]

    Sparse optimizatio n guided pruning for neural networks,

    Y . Shi, A. Tang, L. Niu, and R. Zhou, “Sparse optimizatio n guided pruning for neural networks,” Neurocomputing, vol. 574, p. 127280, 2024

    work page 2024

  24. [24]

    Thinet: A filter level pruni ng method for deep neural network compression,

    J.-H. Luo, J. Wu, and W. Lin, “Thinet: A filter level pruni ng method for deep neural network compression,” in Proceedings of the IEEE international conference on computer vision , 2017, pp. 5058–5066

    work page 2017

  25. [25]

    Soft Filter Pruning for Accelerating Deep Convolutional Neural Networks

    Y . He, G. Kang, X. Dong, Y . Fu, and Y . Y ang, “Soft filter pru ning for accelerating deep convolutional neural networks,” arXiv preprint arXiv:1808.06866, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  26. [26]

    A novel and efficient model pruning method for deep convolutional ne ural networks by evaluating the direct and indirect effects of fil ters,

    Y . Zheng, P . Sun, Q. Ren, W. Xu, and D. Zhu, “A novel and efficient model pruning method for deep convolutional ne ural networks by evaluating the direct and indirect effects of fil ters,” Neurocomputing, vol. 569, p. 127124, 2024. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S092523122301247X

    work page 2024

  27. [27]

    Daar: Dual attention coope rative adaptive pruning rate by data-driven for filter pruning: S. lian et al

    S. Lian, Y . Zhao, and J. Pei, “Daar: Dual attention coope rative adaptive pruning rate by data-driven for filter pruning: S. lian et al. ” Applied Intelligence, vol. 55, no. 6, p. 402, 2025

    work page 2025

  28. [28]

    Eacp: An effec tive automatic channel pruning for neural networks,

    Y . Liu, D. Wu, W. Zhou, K. Fan, and Z. Zhou, “Eacp: An effec tive automatic channel pruning for neural networks,” Neurocomputing, vol. 526, pp. 131–142, 2023

    work page 2023

  29. [29]

    Pro gressive local filter pruning for image retrieval acceleration,

    X. Wang, Z. Zheng, Y . He, F. Y an, Z. Zeng, and Y . Y ang, “Pro gressive local filter pruning for image retrieval acceleration,” IEEE Transactions on Multimedia , vol. 25, pp. 9597–9607, 2023

    work page 2023

  30. [30]

    Sfp: Similarity-based filter pruning for deep neural networks,

    G. Li, R. Li, T. Li, C. Shen, X. Zou, J. Wang, C. Wang, and N. Li, “Sfp: Similarity-based filter pruning for deep neural networks,” Information Sciences, vol. 689, p. 121418, 2025

    work page 2025

  31. [31]

    Convo- lutional neural networks using logarithmic data represent a- tion,

    D. Miyashita, E. H. Lee, and B. Murmann, “Convo- lutional neural networks using logarithmic data represent a- tion,” CoRR, vol. abs/1603.01025, 2016. [Online]. Available: http://arxiv.org/abs/1603.01025

    Pith/arXiv arXiv 2016

  32. [32]

    Incremental n etwork quantization: Towards lossless cnns with low-precision we ights,

    A. Zhou, A. Y ao, Y . Guo, L. Xu, and Y . Chen, “Incremental n etwork quantization: Towards lossless cnns with low-precision we ights,” arXiv preprint arXiv:1702.03044, 2017

    Pith/arXiv arXiv 2017

  33. [33]

    Prom: Prio ritize reduc- tion of multiplications over lower bit-widths for efficient cnns,

    L. Meiner, J. Mehnert, and A. P . Condurache, “Prom: Prio ritize reduc- tion of multiplications over lower bit-widths for efficient cnns,” arXiv preprint arXiv:2505.03254, 2025

    Pith/arXiv arXiv 2025

  34. [34]

    Ql lm: Accurate and efficient low-bitwidth quantization for large language models,

    J. Liu, R. Gong, X. Wei, Z. Dong, J. Cai, and B. Zhuang, “Ql lm: Accurate and efficient low-bitwidth quantization for large language models,” 2024. [Online]. Available: https://arxiv.org/a bs/2310.08041

    Pith/arXiv arXiv 2024

  35. [35]

    Block and subword-s caling floating-point (BSFP) : An efficient non-uniform quantizati on for low precision inference,

    Y .-C. Lo, T.-K. Lee, and R.-S. Liu, “Block and subword-s caling floating-point (BSFP) : An efficient non-uniform quantizati on for low precision inference,” in The Eleventh International Conference on Learning Representations , 2023. [Online]. Available: https://openreview.net/forum?id=VWm4o4l3V9e

    work page 2023

  36. [36]

    Pqk: model compression vi a pruning, quantization, and knowledge distillation,

    J. Kim, S. Chang, and N. Kwak, “Pqk: model compression vi a pruning, quantization, and knowledge distillation,” arXiv preprint arXiv:2106.14681, 2021

    Pith/arXiv arXiv 2021

  37. [37]

    Compressed neural ar- chitecture utilizing dimensionality reduction and quanti zation,

    M. S. Hasan, R. Alam, and M. A. Adnan, “Compressed neural ar- chitecture utilizing dimensionality reduction and quanti zation,” Applied Intelligence, vol. 53, no. 2, pp. 1271–1286, 2023

    work page 2023

  38. [38]

    Quantization-aware trainin g with dynamic and static pruning,

    S. An, J. Shin, and J. Kim, “Quantization-aware trainin g with dynamic and static pruning,” IEEE Access , vol. 13, pp. 57 476–57 484, 2025

    work page 2025

  39. [39]

    Non-structured dnn weight pruning—is it beneficial in any platform?

    X. Ma, S. Lin, S. Y e, Z. He, L. Zhang, G. Y uan, S. H. Tan, Z. L i, D. Fan, X. Qian et al. , “Non-structured dnn weight pruning—is it beneficial in any platform?” IEEE transactions on neural networks and learning systems, vol. 33, no. 9, pp. 4930–4944, 2021

    work page 2021

  40. [40]

    Deep residual learni ng for image recognition,

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

    2016

  41. [41]

    V ery deep convolutional networks for large-scale image recognition,

    K. Simonyan and A. Zisserman, “V ery deep convolutional networks for large-scale image recognition,” arXiv preprint arXiv:1409.1556 , 2014

    Pith/arXiv arXiv 2014

  42. [42]

    Learning multiple layers of features from tiny images,

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

    2009

  43. [43]

    Blen ded coarse gradient descent for full quantization of deep neural netwo rks,

    P . Yin, S. Zhang, J. Lyu, S. Osher, Y . Qi, and J. Xin, “Blen ded coarse gradient descent for full quantization of deep neural netwo rks,” Research in the Mathematical Sciences , vol. 6, pp. 1–23, 2019

    work page 2019

  44. [44]

    Ro bustness- aware 2-bit quantization with real-time performance for ne ural network,

    X. Li, H. Jiang, R. Zhang, F. Tian, S. Huang, and D. Xu, “Ro bustness- aware 2-bit quantization with real-time performance for ne ural network,” Neurocomputing, vol. 455, pp. 12–22, 2021

    work page 2021

  45. [45]

    Se arching for low-bit weights in quantized neural networks,

    Z. Y ang, Y . Wang, K. Han, C. Xu, C. Xu, D. Tao, and C. Xu, “Se arching for low-bit weights in quantized neural networks,” Advances in neural information processing systems , vol. 33, pp. 4091–4102, 2020

    work page 2020

  46. [46]

    Search w hat you want: Barrier panelty nas for mixed precision quantization ,

    H. Y u, Q. Han, J. Li, J. Shi, G. Cheng, and B. Fan, “Search w hat you want: Barrier panelty nas for mixed precision quantization ,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, A ugust 23–28, 2020, Proceedings, Part IX 16 . Springer, 2020, pp. 1–16

    work page 2020

  47. [47]

    Dynamical channel pruning by conditional accuracy change for deep neural netw orks,

    Z. Chen, T.-B. Xu, C. Du, C.-L. Liu, and H. He, “Dynamical channel pruning by conditional accuracy change for deep neural netw orks,” IEEE transactions on neural networks and learning systems , vol. 32, no. 2, pp. 799–813, 2020

    work page 2020

  48. [48]

    Iterative clus tering pruning for convolutional neural networks,

    J. Chang, Y . Lu, P . Xue, Y . Xu, and Z. Wei, “Iterative clus tering pruning for convolutional neural networks,” Knowledge-Based Systems, vol. 265, p. 110386, 2023

    work page 2023

  49. [49]

    Hessian-aware pruning and optimal neural impl ant,

    S. Y u, Z. Y ao, A. Gholami, Z. Dong, S. Kim, M. W. Mahoney, a nd K. Keutzer, “Hessian-aware pruning and optimal neural impl ant,” in Proceedings of the IEEE/CVF Winter Conference on Applicati ons of Computer Vision, 2022, pp. 3880–3891

    work page 2022

  50. [50]

    Concurrent Training and Layer Pruning of Deep Neural Networks

    V . F. I. Guenter and A. Sideris, “Concurrent training an d layer pruning of deep neural networks,” arXiv preprint arXiv:2406.04549 , 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  51. [51]

    Shallowing deep networks: Layer-w ise pruning based on feature representations,

    S. Chen and Q. Zhao, “Shallowing deep networks: Layer-w ise pruning based on feature representations,” IEEE transactions on pattern analysis and machine intelligence , vol. 41, no. 12, pp. 3048–3056, 2018

    work page 2018

  52. [52]

    Inference- aware convolutional neural network pruning,

    T. Choudhary, V . Mishra, A. Goswami, and J. Sarangapani , “Inference- aware convolutional neural network pruning,” Future Generation Com- puter Systems , vol. 135, pp. 44–56, 2022

    work page 2022