Pith. sign in

REVIEW 3 major objections 5 minor 48 references

Loss-Aware Automatic Selection of Structured Pruning Criteria for Deep Neural Network Acceleration

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read LAASP prunes a network during training by greedily choosing, at each step, the layer and filter-importance criterion that minimize loss on a small data subset, matching or beating state-of-the-art structured pruning at equal FLOPs…

desk verdict A useful pruning-while-training recipe, but the loss-aware selection mechanism at its heart is never validated against the final retrained accuracy it is supposed to predict. read the letter →

arxiv 2506.20152 v1 pith:JFBZJPKA submitted 2025-06-25 cs.CV cs.AIeess.IV

Deepak Ghimire , Kilho Lee , Seong-heum Kim This is my paper · ORCID
classification cs.CVcs.AIeess.IV
keywords structuredpruningfilterloss-awarecriterionselectionpruning-while-trainingvariable-rateFLOPsreductionResNetimageclassification
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

Most structured pruning pipelines train a network to convergence, prune it, and then fine-tune it to recover accuracy. This paper claims that all three stages can be collapsed into a single training run: train briefly, then alternate small pruning steps with short retraining bursts. At every pruning step, the method temporarily removes a fixed FLOPs-sized slice from each candidate layer, ranks filters by four cheap criteria ($\ell^1$-norm, $\ell^2$-norm, Euclidean similarity, cosine similarity), measures the network loss on a small random subset, and keeps the layer-criterion pair that loses least. Because each step is loss-guided, the per-layer pruning rates emerge automatically instead of being set by hand. On CIFAR-10 and ImageNet, the resulting ResNet and VGG networks reportedly match or exceed state-of-the-art pruning methods at 42-60% FLOPs reduction, without needing a pretrained model.

What carries the argument

The engine is a greedy one-step lookahead: for each layer $l$ and each criterion $c$, temporarily prune the $E_s^{(l)}$ lowest-ranked filters, evaluate $\mathrm{Loss}(D' \mid W^{*(lc)})$, and commit to the minimizer. The layer-wise exploration step, $E_s^{(l)} = \max(1, P_s \Psi(W) N^{l+1} / (\Psi(W) - \Psi(\bar{W}^{(l)})))$, converts a fixed per-iteration FLOPs budget into an integer filter count so that candidates across layers are comparable; the criterion pool $C = \{ \ell^1\text{-norm}, \ell^2\text{-norm}, \text{euclidean similarity}, \text{cosine similarity} \}$ supplies the filter rankings; and two guards keep the process stable—a per-layer maximum pruning limit $R^{(l)}_{\max}$ in $[0.55, 0.75]$ and a fine-tuning burst of one to three epochs after every $\delta P_{ft} = 3\%$ FLOPs reduction. These pieces are what make per-layer pruning rates emerge automatically rather than being assigned by hand.

What would settle it

Take a fixed model and target FLOPs reduction, and replace Algorithm 1's loss-based argmin over layer–criterion pairs with a random choice among the same candidates while keeping everything else (exploration step, fine-tuning schedule, training budget) identical. If the random selector matches LAASP's accuracy across the same FLOPs reductions, then the immediate post-prune loss is not actually steering the result. A cheaper check: for a set of candidate prunes, compute the rank correlation between $\mathrm{Loss}(D' \mid W^{*(lc)})$ immediately after pruning and the final accuracy after retraining to convergence; low correlation would break the load-bearing premise.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that the two decisions that normally require human judgment in filter pruning—which layer to prune and which importance criterion to trust—can both be made automatically from the network's loss. LAASP first runs a short warm-up (empirically, up to just before the first learning-rate decay), then iteratively prunes the partially trained network. In each iteration it computes, for every convolutional layer, the number of filters $E_s^{(l)}$ whose removal cuts total FLOPs by a fixed ~1% step; ranks filters in that layer under each of four criteria ($\ell^1$-norm, $\ell^2$-norm, Euclidean similarity, cosine similarity); temporarily prunes the lowest-ranked filters; evaluates the candidate subnetwork on a small random subset $D'$ of the training data; and permanently applies the single layer–criterion combination that gives the smallest loss. After every 3% FLOPs reduction the network is fine-tuned for one to three epochs, and after the target rate is reached it is trained to convergence. The reported results are that ResNet56 and ResNet110 on CIFAR-10 retain 93.49% and 94.17% top-1 accuracy at about 52% FLOPs reduction, and ResNet50 on ImageNet loses 0.33% top-5 at 42.3% FLOPs reduction, with no pretrained model.

Load-bearing premise

The load-bearing premise is that the loss measured on a small random subset of training data immediately after a one-step temporary prune ranks the candidates in the same order as their accuracy after subsequent retraining to convergence.

Editorial extensions

If this is right

  • Structured pruning can be folded into normal training: with no separate pretraining phase, a pruned model is obtained within roughly the same training schedule, and the saved computation of running on the slimmed network starts immediately.
  • Per-layer pruning rates no longer need manual allocation; the loss signal decides how many filters each layer can lose, subject only to a shared cap $R^{(l)}_{\max}$.
  • The best criterion varies by architecture—cosine similarity dominates for ResNet110, l1-norm for ResNet56—so fixing one criterion in advance is suboptimal; the loss-guided pool avoids that choice.
  • Because the constraint $\Psi$ is generic, the same selection procedure applies to latency or memory budgets, not just FLOPs.
  • Widening the criterion pool improves accuracy at linearly increasing search cost, and shrinking it to magnitude-only or similarity-only degrades results.

Reading between the lines

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

  • If the immediate-loss ranking is trustworthy, the same greedy rule should transfer to other discrete compression decisions—transformer head removal, layer dropping, or quantization bit assignment—wherever a cost-to-unit conversion like Eq. (8) can be defined; that is an extension the paper does not test.
  • The ablation results suggest the method could be made fully automatic by replacing the fixed pruning epoch with a learned early-pruning indicator, a direction the paper names as future work.
  • A practical deployment check would be to measure how the quality of the selected layer–criterion pairs varies with the size and sampling of $D'$; the paper leaves this as an open operational detail.
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

3 major / 5 minor

Summary. The paper proposes LAASP, a structured (filter) pruning method that interleaves pruning with training. After a short initial training phase, the algorithm iteratively selects, from all prunable layers, the layer and the filter-importance criterion (from l1-norm, l2-norm, Euclidean similarity, and cosine similarity) whose one-step temporary pruning produces the lowest loss on a small random subset of the training data. The chosen filters are permanently removed, the network is fine-tuned periodically during pruning, and training then continues until convergence. The authors claim that this removes the need for a separate pretraining phase, automatically determines per-layer pruning rates and the best criterion, and achieves state-of-the-art results for VGG and ResNet on CIFAR-10 and ImageNet, including 0.33% top-5 drop for ResNet50 at 42.3% FLOPs reduction. Ablations cover the pruning start epoch, the pool of criteria, the maximum per-layer pruning rate, and the exploration step size.

Significance. If the reported results are reproducible, the contribution is practically useful: it simplifies the prune-train pipeline and provides a data-driven way to switch between standard filter-importance criteria. The paper is honest about the heuristic nature of the method and includes code and several ablation studies. However, the evidence presented does not isolate the loss-aware selection mechanism (the defining contribution), and several state-of-the-art comparisons use baselines that are not matched in accuracy, so the strength of the empirical claims is currently not fully established. The headline ImageNet result is nonetheless impressive and warrants a revision that closes these gaps.

major comments (3)
  1. [Algorithm 1, Line 11; Section 4.1] The capping condition in Algorithm 1, Line 11, is inconsistent with the definition of R_max. The text states that R_max is a maximum pruning fraction (e.g., 0.7 means at most 70% of filters can be pruned), but the condition `N'_out >= N_out * R_max` stops pruning a layer as soon as fewer than R_max fraction of the filters remain. With R_max = 0.7 this permits only about 30% pruning, not 70%. A larger R_max would therefore reduce the allowed pruning, contradicting Section 4.1 (where a larger R_max is selected for higher pruning rates) and the ablation in Table 5. Please correct this condition (e.g., to `N'_out > N_out * (1 - R_max)`) and verify that the reported experiments used the corrected version.
  2. [Section 3.5, Algorithm 1 Line 18; Section 4.5.2] The central mechanism of LAASP is the greedy selection of layer and criterion by minimizing one-step post-pruning loss on D' (Line 18). The paper never validates that this one-step loss ordering predicts the validation accuracy after the subsequent fine-tuning and convergence training. Table 4 only compares pools of criteria; it does not compare the greedy loss-based selection with random layer/criterion selection or with any fixed criterion under the same pruning schedule. As a result, the claimed advantage of being 'loss-aware' is not isolated from the pruning-while-training schedule and fine-tuning intervals. In addition, Section 3.5 says only that D' is a 'small, randomly sampled' subset, and Section 4.1 gives no size or sampling protocol; this affects both the reliability of the proxy signal and the reproducibility of the method. I ask for an ablation that fixes the entire pruning pipeline and varies only the selection mechanism (greedy loss vs. random vs. each fixed criterion), plus a report of the D' size and a correlation measurement between the one-step loss and the post-convergence accuracy.
  3. [Tables 2 and 3, Sections 4.3 and 4.4] The state-of-the-art comparison is not controlled for baseline strength. For example, in Table 2 the ResNet110 LAASP baseline is 94.41%, while FPGM, MFP, and MSVFP report baselines of 93.68–93.77%; the pruned LAASP accuracy (94.17%) therefore cannot be directly compared with the other pruned accuracies. Similarly, in Table 3 the ResNet50 LAASP top-1 baseline is 76.48% versus 76.15% for SFP/FPGM/MFP, and the top-1 baseline advantage of 0.33 percentage points is as large as the claimed 0.33% top-5 drop. Since the metric is accuracy drop relative to the method's own baseline, a stronger baseline can make a pruning result look better. Please provide controlled comparisons (same training recipe, same random seeds, or direct re-running of baselines with the LAASP pipeline), or explicitly rescale the drops in the tables.
minor comments (5)
  1. [Section 4.3, first sentence] The text says 'Table 2 presents the efficiency of the pruning of ResNet34, ResNet56, and ResNet110', but the table and the rest of the section refer to ResNet32; correct the model name.
  2. [Section 3.3.2, Eq. (5)] The heading and Eq. (5) use 'Cosign similarity'; this should be 'Cosine similarity'.
  3. [Section 4.1 vs Algorithm 1] The fine-tune FLOPs-reduction interval is denoted δP_ft in Algorithm 1 but βP in Section 4.1; unify the notation.
  4. [Section 4.4 and reference list] The text says 'the PaT method, as described in the paper by Shen et al. [49]', but reference [49] is FuPruner (Li et al.), while Shen et al. is [38]; correct this citation mismatch.
  5. [Section 4.1 and Algorithm 1] The number of initial training epochs tp is not reported for the ResNet CIFAR-10 and ImageNet experiments; please state the exact epoch numbers and the learning-rate schedule so that the pruning-start time is reproducible.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: LAASP's loss-guided greedy selection is a training heuristic evaluated against external accuracy benchmarks, and its central claims do not reduce to their inputs by construction.

full rationale

The central mechanism of LAASP (Algorithm 1, Line 18) selects the layer and filter-importance criterion that minimize the current network loss on a small sampled subset D' after each temporary prune. This is a greedy search heuristic, not a derivation: the final reported accuracies (Tables 1-3) come from external validation on CIFAR-10 and ImageNet, so the headline results are not defined in terms of the selection signal. The paper's only notable self-citation is [26] (Ghimire and Kim, MSVFP), cited in Table 2 as a pretrained baseline comparison and in Related Works as an example of combining magnitude and similarity criteria; that citation is not used to justify the correctness of the proposed pruning-while-training approach and is not load-bearing for any theorem or equation. The known non-circularity weaknesses (D' is unspecified in size, and the one-step loss is not directly validated against post-retraining accuracy) are empirical-validity concerns about whether the greedy signal predicts final accuracy, not cases where an output equals an input by construction. Consequently, the circularity score is 0.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on heuristic importance criteria, an unvalidated loss-proxy assumption, precomputed layer exploration steps, and several empirically tuned hyperparameters. No fundamentally new physical entities are introduced.

free parameters (6)
  • R_max (layer-wise maximum pruning rate) = 0.55 to 0.75, chosen per target FLOPs
    Ablation on ResNet20/CIFAR-10 (Table 5) supports a range, but the value per experiment is hand/empirically chosen and caps per-layer pruning, directly affecting the final layer-wise pruning rates.
  • Ps (exploration step FLOPs reduction rate) = 0.01
    Selected from ablation on ResNet20/CIFAR-10 (Table 6) as a trade-off between accuracy and complexity.
  • delta_P_ft (fine-tune FLOPs reduction interval) = 0.03
    Set to 3% FLOPs reduction between fine-tune sessions; no dedicated ablation is reported.
  • epoch_ft (fine-tune epochs per interval) = 1 to 3
    Heuristic range stated in Section 4.1; not ablated in the paper.
  • tp (training epochs before pruning) = Not stated per dataset; 'just before initial LR decay'
    Chosen based on Fig. 7 (ResNet20/CIFAR-10) and stated empirically; exact epoch numbers are not in the paper.
  • Size of loss-estimation subset D' = Not reported
    Critical to the greedy selection; its size and sampling procedure are not specified, making the loss estimates irreproducible and their reliability unverifiable.
assumptions (5)
  • domain assumption Filter magnitude (l1/l2 norm) and filter redundancy (Euclidean/cosine similarity) are informative proxies for which filters can be removed.
    Invoked in Section 3.3 to define the candidate criteria pool; these are heuristic importance measures, not derived from the pruning objective.
  • domain assumption The loss on a small random subset D' after a one-step temporary pruning ranks candidate pruned networks by their eventual post-retraining accuracy.
    Core to Algorithm 1, Line 18 and Section 3.5; unvalidated and subset size unspecified.
  • domain assumption The layer-wise exploration step computed once from the original network (Eq. 8) remains a valid FLOPs-reduction estimate throughout pruning.
    Section 3.4 states the step is calculated on the original network and maintained, but FLOPs per layer change as filters are removed.
  • ad hoc to paper Pruning just before the initial learning-rate decay yields the best accuracy under a fixed total training budget.
    Established empirically in Section 4.5.1 (Fig. 7) on ResNet20/CIFAR-10; no general theoretical justification.
  • domain assumption Standard structured pruning bookkeeping: removing a filter also removes corresponding input channels in downstream layers.
    Implicit in the hard-pruning strategy and FLOPs accounting; conventional in the field.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Loss-Aware Automatic Selection of Structured Pruning Criteria for Deep Neural Network Acceleration." pith.science (2026). https://pith.science/paper/JFBZJPKA

@misc{pith2026250620152,
  author       = {Pith},
  title        = {Pith review of: Loss-Aware Automatic Selection of Structured Pruning Criteria for Deep Neural Network Acceleration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JFBZJPKA}},
  note         = {Machine review of arXiv:2506.20152}
}
read the original abstract

Structured pruning is a well-established technique for compressing neural networks, making it suitable for deployment in resource-limited edge devices. This paper presents an efficient Loss-Aware Automatic Selection of Structured Pruning Criteria (LAASP) for slimming and accelerating deep neural networks. The majority of pruning methodologies employ a sequential process consisting of three stages: 1) training, 2) pruning, and 3) fine-tuning, whereas the proposed pruning technique adopts a pruning-while-training approach that eliminates the first stage and integrates the second and third stages into a single cycle. The automatic selection of magnitude or similarity-based filter pruning criteria from a specified pool of criteria and the specific pruning layer at each pruning iteration is guided by the network's overall loss on a small subset of the training data. To mitigate the abrupt accuracy drop due to pruning, the network is retrained briefly after each reduction of a predefined number of floating-point operations (FLOPs). The optimal pruning rates for each layer in the network are automatically determined, eliminating the need for manual allocation of fixed or variable pruning rates for each layer. Experiments on the VGGNet and ResNet models on the CIFAR-10 and ImageNet benchmark datasets demonstrate the effectiveness of the proposed method. In particular, the ResNet56 and ResNet110 models on the CIFAR-10 dataset significantly improve the top-1 accuracy compared to state-of-the-art methods while reducing the network FLOPs by 52\%. Furthermore, the ResNet50 model on the ImageNet dataset reduces FLOPs by more than 42\% with a negligible 0.33\% drop in top-5 accuracy. The source code of this paper is publicly available online - https://github.com/ghimiredhikura/laasp.

Figures

Figures reproduced from arXiv: 2506.20152 by the authors.

Figure 1
Figure 1. An illustration of the proposed model training and pruning process. Initially, the original model is [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Step pruning rate for each layer of the ResNet18 model evaluated in ImageNet dataset. The overall [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. The distribution of the total number of pruned filters for ResNet32, ResNet56, and ResNet110 on [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: The number of pruned and retained filters in each layer of ResNet32 after pruning on the CIFAR-10 [PITH_FULL_IMAGE:figures/full_fig_p019_4.png]
Figure 5
Figure 5. Figure 5: The distribution of the total number of pruned filters for ResNet18, ResNet34, and ResNet50 on [PITH_FULL_IMAGE:figures/full_fig_p021_5.png]
Figure 6
Figure 6. Figure 6: The number of pruned and retained filters in each layer of ResNet34 after pruning on the ImageNet [PITH_FULL_IMAGE:figures/full_fig_p022_6.png]
Figure 7
Figure 7. Figure 7: Top-1 validation accuracy after pruning ResNet20 model for CIFAR-10 dataset at different network [PITH_FULL_IMAGE:figures/full_fig_p023_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 36 canonical work pages

  1. [1]

    Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding,

    S. Han, H. Mao, and W. J. Dally, “Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding,”arXiv preprint arXiv:1510.00149, 2015

    arXiv 2015

  2. [2]

    Learning both weights and connections for efficient neural network,

    S. Han, J. Pool, J. Tran, and W. Dally, “Learning both weights and connections for efficient neural network,”Advances in neural information processing systems, vol. 28, 2015

    2015

  3. [3]

    The lottery ticket hypothesis: Finding sparse, trainable neural networks,

    J. Frankle and M. Carbin, “The lottery ticket hypothesis: Finding sparse, trainable neural networks,” arXiv preprint arXiv:1803.03635, 2018

    arXiv 2018

  4. [4]

    Global sparse momentum sgd for pruning very deep neural networks,

    X. Ding, X. Zhou, Y . Guo, J. Han, J. Liu,et al., “Global sparse momentum sgd for pruning very deep neural networks,” Advances in Neural Information Processing Systems, vol. 32, 2019

    work page 2019

  5. [5]

    Soft filter pruning for accelerating deep convolutional neural networks,

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

    arXiv 2018

  6. [6]

    Learning filter pruning criteria for deep convolutional neural networks acceleration,

    Y . He, Y . Ding, P. Liu, L. Zhu, H. Zhang, and Y . Yang, “Learning filter pruning criteria for deep convolutional neural networks acceleration,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 2009–2018

    work page 2020

  7. [7]

    Filter pruning via geometric me- dian for deep convolutional neural networks acceleration,

    Y . He, P. Liu, Z. Wang, Z. Hu, and Y . Yang, “Filter pruning via geometric me- dian for deep convolutional neural networks acceleration,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2019, pp. 4340–4349

    work page 2019

  8. [8]

    Filter pruning by switching to neighboring cnns with good attributes,

    Y . He, P. Liu, L. Zhu, and Y . Yang, “Filter pruning by switching to neighboring cnns with good attributes,”IEEE Transactions on Neural Networks and Learning Systems, 2022

    work page 2022

Show all 48 references
  1. [10]

    Post training 4-bit quantization of con- volutional networks for rapid-deployment,

    R. Banner, Y . Nahshan, and D. Soudry, “Post training 4-bit quantization of con- volutional networks for rapid-deployment,”Advances in Neural Information Pro- cessing Systems, vol. 32, 2019

    2019

  2. [11]

    Quantization and training of neural networks for efficient integer-arithmetic-only inference,

    B. Jacob, S. Kligys, B. Chen, M. Zhu, M. Tang, A. Howard, H. Adam, and D. Kalenichenko, “Quantization and training of neural networks for efficient integer-arithmetic-only inference,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 2704–2713

    2018

  3. [12]

    Model compression via distillation and quantization,

    A. Polino, R. Pascanu, and D. Alistarh, “Model compression via distillation and quantization,” arXiv preprint arXiv:1802.05668, 2018

    2018 arXiv

  4. [13]

    Refine myself by teaching myself: Feature refinement via self-knowledge distillation,

    M. Ji, S. Shin, S. Hwang, G. Park, and I.-C. Moon, “Refine myself by teaching myself: Feature refinement via self-knowledge distillation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2021, pp. 10 664–10 673

    2021

  5. [14]

    Learning low-rank deep neural networks via singular vector orthogonality regularization and singular value sparsification,

    H. Yang, M. Tang, W. Wen, F. Yan, D. Hu, A. Li, H. Li, and Y . Chen, “Learning low-rank deep neural networks via singular vector orthogonality regularization and singular value sparsification,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognitio...

    2020

  6. [15]

    Towards efficient tensor decomposition- based dnn model compression with optimization framework,

    M. Yin, Y . Sui, S. Liao, and B. Yuan, “Towards efficient tensor decomposition- based dnn model compression with optimization framework,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2021, pp. 10 674–10 683

    2021

  7. [16]

    Supernet in neural architecture search: A taxonomic survey,

    S. Cha, T. Kim, H. Lee, and S.-Y . Yun, “Supernet in neural architecture search: A taxonomic survey,”arXiv preprint arXiv:2204.03916, 2022

    2022 arXiv

  8. [18]

    Model compression and hardware acceleration for neural networks: A comprehensive survey,

    L. Deng, G. Li, S. Han, L. Shi, and Y . Xie, “Model compression and hardware acceleration for neural networks: A comprehensive survey,” Proceedings of the IEEE, vol. 108, no. 4, pp. 485–532, 2020

    2020

  9. [19]

    A survey on efficient convolutional neural networks and hardware acceleration,

    D. Ghimire, D. Kil, and S.-h. Kim, “A survey on efficient convolutional neural networks and hardware acceleration,” Electronics, vol. 11, no. 6, p. 945, 2022

    2022

  10. [20]

    Pruning filters for efficient convnets,

    H. Li, A. Kadav, I. Durdanovic, H. Samet, and H. P. Graf, “Pruning filters for efficient convnets,”arXiv preprint arXiv:1608.08710, 2016

    2016 arXiv

  11. [21]

    Thinet: A filter level pruning method for deep neu- ral network compression,

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

    2017

  12. [22]

    Nisp: Pruning networks using neuron importance score prop- agation,

    R. Yu, A. Li, C.-F. Chen, J.-H. Lai, V . I. Morariu, X. Han, M. Gao, C.-Y . Lin, and L. S. Davis, “Nisp: Pruning networks using neuron importance score prop- agation,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 9194–9203

    2018

  13. [23]

    Automated filter pruning based on high- dimensional bayesian optimization,

    T. Kim, H. Choi, and Y . Choe, “Automated filter pruning based on high- dimensional bayesian optimization,” IEEE Access, vol. 10, pp. 22 547–22 555, 2022

    2022

  14. [24]

    Adaptive cnn filter pruning using global importance metric,

    M. Mondal, B. Das, S. D. Roy, P. Singh, B. Lall, and S. D. Joshi, “Adaptive cnn filter pruning using global importance metric,” Computer Vision and Image Understanding, p. 103511, 2022

    2022

  15. [25]

    Filter pruning without damaging networks capacity,

    Y . Zuo, B. Chen, T. Shi, and M. Sun, “Filter pruning without damaging networks capacity,”IEEE Access, vol. 8, pp. 90 924–90 930, 2020

    2020

  16. [26]

    Magnitude and similarity based variable rate fil- ter pruning for efficient convolution neural networks,

    D. Ghimire and S.-H. Kim, “Magnitude and similarity based variable rate fil- ter pruning for efficient convolution neural networks,” Applied Sciences, vol. 13, no. 1, p. 316, 2023

    2023

  17. [27]

    Depgraph: Towards any structural pruning,

    G. Fang, X. Ma, M. Song, M. B. Mi, and X. Wang, “Depgraph: Towards any structural pruning,” arXiv preprint arXiv:2301.12900, 2023. © 2023. This manuscript version is made available under the CC-BY-NC-ND 4.0 license. Published version: https://doi.org/10.1016/j.imavis.2023.104745

    2023 arXiv

  18. [28]

    On the channel pruning using graph convolution network for convolutional neural network acceleration,

    Y . C. Di Jiang and Q. Yang, “On the channel pruning using graph convolution network for convolutional neural network acceleration,” in Proc. Int. Joint Conf. Artif. Intell, vol. 7, 2022, pp. 3107–3113

    2022

  19. [29]

    Optimiz- ing deep neural networks on intelligent edge accelerators via flexible-rate filter pruning,

    G. Li, X. Ma, X. Wang, H. Yue, J. Li, L. Liu, X. Feng, and J. Xue, “Optimiz- ing deep neural networks on intelligent edge accelerators via flexible-rate filter pruning,” Journal of Systems Architecture, vol. 124, p. 102431, 2022

    2022

  20. [30]

    Falf convnets: Fatuous auxiliary loss based filter-pruning for efficient deep cnns,

    P. Singh, V . S. R. Kadi, and V . P. Namboodiri, “Falf convnets: Fatuous auxiliary loss based filter-pruning for efficient deep cnns,” Image and Vision Computing , vol. 93, p. 103857, 2020

    2020

  21. [31]

    Performance-aware approxima- tion of global channel pruning for multitask cnns,

    H. Ye, B. Zhang, T. Chen, J. Fan, and B. Wang, “Performance-aware approxima- tion of global channel pruning for multitask cnns,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023

    2023

  22. [32]

    Pruning neural networks at initialization: Why are we missing the mark?

    J. Frankle, G. K. Dziugaite, D. Roy, and M. Carbin, “Pruning neural networks at initialization: Why are we missing the mark?” in International Conference on Learning Representations , 2021. [Online]. Available: https: //openreview.net/forum?id=Ig-VyQc-MLK

    2021

  23. [33]

    Linear mode connectivity and the lottery ticket hypothesis,

    J. Frankle, G.-K. Dziugaite, D. Roy, and M. Carbin, “Linear mode connectivity and the lottery ticket hypothesis,” in Proceedings of the 37th International Conference on Machine Learning , ser. Proceedings of Machine Learning Research, H. D. III and A. Singh, Eds., vol. 119. PM...

    2020

  24. [34]

    Prune- train: fast neural network training by dynamic sparse model reconfiguration,

    S. Lym, E. Choukse, S. Zangeneh, W. Wen, S. Sanghavi, and M. Erez, “Prune- train: fast neural network training by dynamic sparse model reconfiguration,” in Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, 2019, pp. 1–13

    2019

  25. [35]

    Oyedotun, D

    O. Oyedotun, D. Aouada, and B. Ottersten, “Structured compression of deep neu- © 2023. This manuscript version is made available under the CC-BY-NC-ND 4.0 license. Published version: https://doi.org/10.1016/j.imavis.2023.104745 ral networks with debiased elastic group lasso,” ...

    2023

  26. [36]

    Only train once: A one-shot neural network training and pruning framework,

    T. Chen, B. Ji, T. Ding, B. Fang, G. Wang, Z. Zhu, L. Liang, Y . Shi, S. Yi, and X. Tu, “Only train once: A one-shot neural network training and pruning framework,” in Advances in Neural Information Processing Systems, M. Ranzato, A. Beygelzimer, Y . Dauphin, P. Liang, and J. ...

    2021

  27. [37]

    Otov2: Automatic, generic, user-friendly,

    T. Chen, L. Liang, T. Ding, Z. Zhu, and I. Zharkov, “Otov2: Automatic, generic, user-friendly,”arXiv preprint arXiv:2303.06862, 2023

    2023 arXiv

  28. [38]

    When to prune? a policy towards early structural pruning,

    M. Shen, P. Molchanov, H. Yin, and J. M. Alvarez, “When to prune? a policy towards early structural pruning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 12 247–12 256

    2022

  29. [39]

    Learning multiple layers of features from tiny images,

    A. Krizhevsky, “Learning multiple layers of features from tiny images,” pp. 32–33, 2009. [Online]. Available: https://www.cs.toronto.edu/ ∼kriz/ learning-features-2009-TR.pdf

    2009

  30. [40]

    Imagenet large scale visual recog- nition challenge,

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

    2015

  31. [41]

    Deep residual learning for image recog- nition,

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

    2016

  32. [42]

    Very deep convolutional networks for large- scale image recognition,

    K. Simonyan and A. Zisserman, “Very deep convolutional networks for large- scale image recognition,” arXiv preprint arXiv:1409.1556, 2014. © 2023. This manuscript version is made available under the CC-BY-NC-ND 4.0 license. Published version: https://doi.org/10.1016/j.imavis.2...

    2014 arXiv

  33. [43]

    Identity mappings in deep residual net- works,

    K. He, X. Zhang, S. Ren, and J. Sun, “Identity mappings in deep residual net- works,” in European conference on computer vision. Springer, 2016, pp. 630– 645

    2016

  34. [44]

    Wide residual networks,

    S. Zagoruyko and N. Komodakis, “Wide residual networks,” arXiv preprint arXiv:1605.07146, 2016

    2016 arXiv

  35. [45]

    Automatic differentiation in pytorch,

    A. Paszke, S. Gross, S. Chintala, G. Chanan, E. Yang, Z. DeVito, Z. Lin, A. Des- maison, L. Antiga, and A. Lerer, “Automatic differentiation in pytorch,” 2017

    2017

  36. [46]

    Hrank: Filter pruning using high-rank feature map,

    M. Lin, R. Ji, Y . Wang, Y . Zhang, B. Zhang, Y . Tian, and L. Shao, “Hrank: Filter pruning using high-rank feature map,” in Proceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition, 2020, pp. 1529–1538

    2020

  37. [47]

    Network pruning via performance maximization,

    S. Gao, F. Huang, W. Cai, and H. Huang, “Network pruning via performance maximization,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 9270–9280

    2021

  38. [48]

    Rethinking the value of network pruning,

    Z. Liu, M. Sun, T. Zhou, G. Huang, and T. Darrell, “Rethinking the value of network pruning,” arXiv preprint arXiv:1810.05270, 2018

    2018 arXiv

  39. [49]

    Fusion-catalyzed pruning for optimizing deep learning on intelligent edge devices,

    G. Li, X. Ma, X. Wang, L. Liu, J. Xue, and X. Feng, “Fusion-catalyzed pruning for optimizing deep learning on intelligent edge devices,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems , vol. 39, no. 11, pp. 3614–3626, 2020

    2020

  40. [50]

    Leveraging filter correlations for deep model compression,

    P. Singh, V . K. Verma, P. Rai, and V . Namboodiri, “Leveraging filter correlations for deep model compression,” in Proceedings of the IEEE/CVF Winter Confer- ence on applications of computer vision, 2020, pp. 835–844

    2020

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.