Optimal Brain Connection: Towards Efficient Structural Pruning

Binhua Huang; Deepu John; Guoxin Wang; Lei Huang; Qingyuan Wang; Shaowu Chen; Wei Ma; Weize Sun

REVIEW 3 major objections 6 minor 2 cited by

Optimal Brain Connection: Towards Efficient Structural Pruning

T0 review · 3 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The paper claims that structural pruning should be guided by a Jacobian-based saliency score that sums squared loss-gradient interactions over linked parameter groups, plus an autoencoder fine-tuning step that keeps pruned connections alive

desk verdict A solid, well-tested pruning method with an overclaimed interaction story; the Jacobian criterion is a useful first-order metric but does not capture the cross-layer dependencies it advertises. read the letter →

arxiv 2508.05521 v1 pith:Z32YH2Y7 submitted 2025-08-07 cs.CV

Shaowu Chen , Wei Ma , Binhua Huang , Qingyuan Wang , Guoxin Wang , Weize Sun , Lei Huang , Deepu John This is my paper

classification cs.CV

keywords structuralpruningJacobianCriterionlosssaliencyEquivalentautoencoderchannelmodelcompressionVisionTransformer

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 is trying to establish a better answer to a basic pruning question: which filters or channels can be removed with the least damage? It proposes the Jacobian Criterion, a first-order score $S(\{w_m \mid m \in G_g\}) = \sum_{m \in G_g} w_m^\top J_m^\top J_m w_m$, which measures how much removing a coupled structural group would perturb the loss vector, and it argues this catches interactions that per-weight Taylor or Fisher scores miss. It also introduces Equivalent Pruning, a compressor-decompressor inserted around the layer during fine-tuning so that even connections scheduled for removal keep contributing to recalibration, and then merged so the final architecture is identical to naive pruning. If true, the paper offers a cheap but accurate ranking metric plus a training-time trick that should make structural pruning safer at high compression rates across CNNs, ViTs, object detectors, and LLMs.

What carries the argument

The load-bearing object is the Jacobian Criterion, Eq. (8), a structural saliency score that sums per-parameter quadratic forms $w_m^\top J_m^\top J_m w_m$ over a coupled group, where $J_m$ is the Jacobian of the per-batch loss vector with respect to that parameter. It is derived from the first-order expansion of the loss vector combined with a block-diagonal approximation of $J^\top J$, which makes the full matrix computation unnecessary. The second mechanism is Equivalent Pruning: paired linear layers $C$ (compressor) and $D$ (decompressor) inserted around a pruned layer as 1x1 convolutions or linear layers, initialized from the identity with pruned rows removed, fine-tuned, then multiplie

What would settle it

Choose ResNet-56 on CIFAR-10, compute the full $J^\top J$ (or a sampling of off-diagonal blocks) for two structural groups with nearly equal JC scores, prune each group alone from the pretrained model, and measure the actual validation loss change. If the group judged less important by JC consistently causes the larger loss increase, the criterion's ordering is falsified; likewise, if the empirical loss perturbation of pruning two groups together is far from the sum of their individual JC scores, the block-diagonal approximation breaks.

Watch

Extended reading notes

Core claim

The core discovery is that the squared loss perturbation caused by pruning can be approximated by a quadratic form in the loss-gradient Jacobian, $L(\Delta w) = \Delta w^\top J^\top J \Delta w$, and that for structural pruning the right saliency of a group is Eq. (8): $S(\{w_m \mid m \in G_g\}) = \sum_{m\in G_g} w_m^\top J_m^\top J_m w_m$. Unlike Taylor importance, which effectively uses a diagonal mask $I \odot J^\top J$, or Fisher's Hessian approximation, which relies on a zero-gradient assumption, JC keeps the off-diagonal blocks within each structural parameter and sums coupled parameters (filter weights, BN scale/shift, and the corresponding incoming channels of the next layer). The pap

Load-bearing premise

The ranking assumes that structural groups do not interact: the damage from removing one filter or channel is independent of the damage from removing another, so their saliency scores can simply be added.

Editorial extensions

If this is right

If JC ranks groups accurately, global pruning without per-layer normalization becomes safe, avoiding the layer-collapse failures that normalization-based methods show at high compression rates.
If EP works as claimed, pruned models can be fine-tuned with the same final architecture as naive pruning while retaining the original connections' information, so recovery is stronger and inference cost is unchanged.
If the reported timings hold, JC is nearly as cheap as Taylor (2.73 s vs 2.66 s per step on ResNet-56) and orders of magnitude cheaper than Hessian-based evaluation (242.80 s), making it deployable where Hessian methods are not.
If the generalization results hold, the same criterion applies to convolutional filters, transformer feed-forward blocks, object-detection backbones, and language-model heads, giving one unified saliency metric across architectures.
If the ImageNet result is representative, OBC can remove more than half of ResNet-50's MACs while slightly improving accuracy (+0.42%), so pruning need not cost accuracy at moderate rates.

Reading between the lines

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

Editorial extension: because JC is a Gauss-Newton-type approximation of the Hessian, a natural next step is to use low-rank off-block corrections to $J^\top J$; that would test whether the block-diagonal assumption, not the quadratic form itself, is the bottleneck at very high pruning rates.
Editorial extension: the EP autoencoder is a trainable analogue of OBS's recalibration step; using it in iterative prune-finetune-merge cycles, or extending the merge to grouped convolutions via reshaping, would likely compound the reported gains.
Editorial extension: the abrupt accuracy collapse in the 'Mean Mean' configuration of Table 7 suggests that when scores are averaged and normalized, the equality between the sum of per-group saliencies and the true loss perturbation degrades; a targeted experiment comparing 'Sum None' and 'Mean Mean' on the same network's full loss perturbation would isolate when the approximation fails.

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.

Desk Editor's Note

read the letter

First, the useful part: the Jacobian Criterion is a reasonably new and clean way to get a per-structural-group saliency from a squared-loss Taylor expansion, keeping the full J^T J within a parameter block rather than diagonalizing it. That matters: Figure 4 shows the off-diagonals are not negligible, and the no-fine-tuning comparisons in Figures 5-7 show it consistently beats Taylor/Fisher/group-norm baselines. The EP autoencoder is a neat reparameterization of soft pruning—retain all weights, learn a compressor/decompressor, then merge—and Table 7 shows it gives a fairly consistent boost in fine-tuned accuracy. Code is released. This is a real empirical contribution.

The soft spots are in the claims, not the experiments. The abstract and methodology say JC 'explicitly captures both intra-component interactions and inter-layer dependencies.' But Eq. (4) assumes J^T J is block-diagonal: off-block correlations are set to zero. Then Eq. (8) aggregates group saliency as a sum of per-parameter terms. So the inter-layer coupling is captured only as an additive sum of marginals, not as an interaction term. The paper never measures the off-block norm or compares against the exact J^T J ranking. The only ablation of interactions (Figure 8) zeros within-block off-diagonals for BN, not off-block blocks. That is a real gap between narrative and derivation.

Table 7 gives a concrete symptom. With the 'Mean Mean' configuration, JC collapses to 35.94% at 6x MACs on VGG19/CIFAR-100 while DepGraph holds 71.47%. The default Sum/None config is fine, but the criterion is fragile to a simple aggregation change in a way that suggests the score is not capturing what the story says it captures.

Also worth flagging: the ImageNet fine-tuning recipe uses distillation (coefficient 0.5 for ResNet-50 and ViT). The comparison methods are not all run under the same distillation recipe. That does not invalidate the results, but it muddies the head-to-head at the top of Table 2. And on MobileNet-v2, Jacobian without EP actually lands below DepGraph (68.12 vs 68.46), which is a useful counterpoint to the blanket claim of superiority.

The math is mostly sound; the squared-loss objective is admittedly an ad hoc proxy, and the derivation leans on the block-diagonal assumption. The limitation statement in the appendix honestly notes EP's overhead and incompatibility with group convolutions, which is good practice.

Bottom line: this is a serious paper with a useful first-order metric and a clever fine-tuning trick. The interaction claim needs to be scaled back or supported with off-block measurements. A referee should ask for that quantification and for a version of Table 2 that isolates distillation. But the default-config results and released code make it worth serious review, and I would tell the authors to fix the overclaim rather than reject.

Referee Report

3 major / 6 minor

Summary. The paper proposes a structural pruning framework, Optimal Brain Connection (OBC), with two components. First, the Jacobian Criterion (JC) estimates the saliency of structural parameter groups by a first-order Taylor expansion of a squared per-batch loss-deviation vector, yielding a quadratic form (Δw)^T J^T J Δw. To make this tractable, J^T J is assumed block-diagonal over structural parameters, and group saliency is computed as a sum of per-parameter quadratic forms. Second, Equivalent Pruning (EP) inserts learnable compressor/decompressor layers (C and D) during fine-tuning, retaining all original connections while simulating the pruned topology, and later merges these layers to recover the naively pruned architecture. Experiments on ImageNet, CIFAR-10/100, ViT-B/16, YOLOv7, and Phi-3-mini report that JC outperforms several data-driven criteria and that EP improves fine-tuned accuracy. Code is released.

Significance. If the claims hold, the paper's main contributions are a cheap first-order saliency metric that captures intra-component (within-filter or within-channel) parameter interactions through the dense J_m^T J_m blocks, and a practical reparameterization that allows pruned parameters to contribute during fine-tuning without changing the final inference architecture. The paper is not circular: the saliency is computed from pre-trained network gradients and no parameter is fit to the final accuracy. The experimental breadth (CNNs, ViT, object detection, LLM) is a clear strength, and releasing code supports reproducibility. However, the central claim that the criterion 'explicitly captures both intra-component interactions and inter-layer dependencies' is only partially supported by the mathematics: intra-component interactions are modeled, but inter-layer dependencies appear only as additive aggregation of per-parameter scores, not as interaction terms. The unexplained catastrophic collapse in Table 7 under a simple configuration change further weakens the robustness claim.

major comments (3)

[Methodology, Eqs. (4), (6), (8)] The paper claims that the Jacobian Criterion 'explicitly captures both intra-component interactions and inter-layer dependencies,' but Eq. (4) assumes J^T J is block-diagonal across the M structural parameters, setting all off-block (including cross-layer) correlations to zero. Eq. (8) then defines group saliency as a sum of per-parameter quadratic forms. For the example group {w_m^(l), b_m^(l), w'_{m+1}^(l+1)}, cross terms such as (w_m^(l))^T J_m^T J_b b_m^(l) are discarded by construction. Thus inter-layer dependencies enter only as additive aggregation, not as interaction terms. This is load-bearing because the claimed advantage over Taylor and Fisher criteria is attributed to capturing parameter connections. I request either (a) a corrected claim, (b) a computation or approximation of the cross-block terms for grouped parameters, or (c) quantitative evidence that the off-block norm o
[Table 7, 'Mean Mean' rows] For VGG-19/CIFAR-100 at 6× MACs, the Jacobian Criterion collapses to 35.94% (and 28.11% at 9×), while DepGraph retains 71.47% (68.40% at 9×); EP does not prevent the collapse (38.08% with EP at 6×). This is a catastrophic failure under a simple choice of aggregator and normalizer. The paper reports the ablation but does not discuss or explain this behavior. This directly qualifies the claim in the Methodology that JC 'achieves effective global pruning without normalization' and the narrative of robustness. Please analyze the cause (e.g., score-scale distortion, rank inversion, cancellation in the additive group score) and either fix the criterion or explicitly delimit the safe configuration space. As it stands, Table 7 shows the method is fragile to a small change in aggregation.
[Methodology, Eqs. (1)-(3)] The saliency is derived from a squared-error proxy L(Δw) = ||l(w+Δw) − l(w)||^2, not the true loss change. The squared form is introduced without justification; the actual scalar loss degradation would be l(w+Δw) − l(w) (or its absolute value). The choice to square is consequential: it produces the quadratic form J^T J and changes the ranking (e.g., it emphasizes samples with large loss deviations). As this proxy is the foundation of the criterion, its validity should be motivated or empirically validated against the true loss change for small perturbations. Without such support, the connection between the score S and actual pruning damage is an assumption rather than a derivation.

minor comments (6)

[Eq. (3)] 'J^T J is positive definite in practice' is generally false when the number of batches N is smaller than the number of parameters P, which is typical (N=50, P in millions). J^T J is then positive semidefinite and may have zero directions; the subsequent statement that any nonzero perturbation increases L is too strong. Please correct to positive semidefinite.
[Figures 5 and 6] Figure 5 reports train accuracy for VGG-19, while Figure 6 reports validation accuracy for ResNet-56 and ResNet-50. The inconsistency in the evaluation metric should be clarified, and ideally the same metric should be used for all comparisons.
[Table 2, MobileNet-v2 row] For MobileNet-v2, Jacobian (68.12) is below DepGraph (68.46) at the same MACs (0.15B). The abstract's 'outperforms several popular metrics' is true but not universal; please add a sentence interpreting this result and note any limitations for inverted-bottleneck architectures.
[Algorithm 1] Line 10 says 'Prune (p·G) groups'; if p·G is not an integer, the behavior is undefined. Please specify rounding or impose a minimum of one group per iteration.
[Eq. (16)] The notation 'I_{O1}[{1,...,O1} ⊘ P, :]' uses '⊘' without definition; presumably it means set difference. Please clarify the indexing notation for the identity-row selection.
[Table 3 and references] The entries labeled 'OBD (Wang et al. 2019)' appear twice with different speedups, but the original OBD is LeCun, Denker, and Solla (1989). The citation attribution is confusing and should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Jacobian Criterion and Equivalent Pruning are derived from stated approximations and trained optimization, not from the target results.

full rationale

The paper's central derivation is self-contained. The Jacobian Criterion is obtained from a first-order Taylor expansion of a squared loss-vector perturbation: Eq. (3) gives L(Δw) = ΔwᵀJᵀJΔw, and Eq. (4) introduces the block-diagonal approximation for JᵀJ as an explicit computational assumption, not as something imported from the paper's own conclusions. Eq. (8) then defines group saliency as a sum of per-parameter quadratic forms S = Σ wₘᵀJₘᵀJₘwₘ. This is a metric computed from network gradients; no parameter is fitted to the reported accuracies, and no prediction is used to set constants. The Equivalent Pruning autoencoder (C and D layers) is a learnable transformation trained during fine-tuning; training an optimization module is not circular derivation. The self-citations (WHC, coupled tensor decomposition) are background comparisons and are not load-bearing for the main claim. The block-diagonal assumption and the Table 7 'Mean Mean' collapse are legitimate correctness/robustness concerns about whether the approximation captures cross-layer interactions, but they are not circularity: the paper does not define its criterion in terms of the results, nor does it rely on a self-citation to force its choice. The derivation chain is therefore not circular, and the honest finding is a score of 0.

Assumptions & free parameters 2 free parameters · 4 assumptions · 1 invented entities

The central claim rests on a Taylor expansion, a block-diagonal approximation, a sum-aggregation for groups, and an ad hoc squared-loss objective. No free parameters are fitted to data to produce the reported accuracy; hyperparameters are standard.

free parameters (2)

Number of gradient batches N = 50 (default)
The saliency estimate uses N=50 data batches for all models; results vary with N (Figure 7b) but the default is chosen for computational cost.
Pruning step size p = 1/400 for CIFAR, 1/100 for ImageNet
Iteratively prunes a fraction of groups per step; smaller p improves greedy approximation but increases cost.

assumptions (4)

standard math Taylor expansion of the loss vector l(w+Δw) ≈ l(w) + JΔw, ignoring higher-order terms
Eq. (2) in Methodology; the central derivation of JC.
domain assumption J^T J is block-diagonal over structural parameters; off-block correlations are zero
Eq. (4); empirically motivated by Figure 3, but not proven.
domain assumption Pruning a group does not change saliencies of other groups, and group saliency is the sum of individual parameter saliencies
Eq. (8); this is the aggregation used, ignoring cross-group and cross-layer interaction terms.
ad hoc to paper The squared loss degradation L(Δw) is a valid proxy for pruning damage
The paper defines L as the sum of squared loss deviations, a non-standard objective chosen to make the criterion positive definite.

invented entities (1)

Compressor C and Decompressor D layers
purpose: Temporary autoencoder inserted around pruned layers to retain all original connections during fine-tuning; merged afterward.
These are architectural constructs, not independently verified entities; their benefit is only shown through accuracy after fine-tuning.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Optimal Brain Connection: Towards Efficient Structural Pruning." pith.science (2026). https://pith.science/paper/Z32YH2Y7

@misc{pith2026250805521,
  author       = {Pith},
  title        = {Pith review of: Optimal Brain Connection: Towards Efficient Structural Pruning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z32YH2Y7}},
  note         = {Machine review of arXiv:2508.05521}
}

read the original abstract

Structural pruning has been widely studied for its effectiveness in compressing neural networks. However, existing methods often neglect the interconnections among parameters. To address this limitation, this paper proposes a structural pruning framework termed Optimal Brain Connection. First, we introduce the Jacobian Criterion, a first-order metric for evaluating the saliency of structural parameters. Unlike existing first-order methods that assess parameters in isolation, our criterion explicitly captures both intra-component interactions and inter-layer dependencies. Second, we propose the Equivalent Pruning mechanism, which utilizes autoencoders to retain the contributions of all original connection--including pruned ones--during fine-tuning. Experimental results demonstrate that the Jacobian Criterion outperforms several popular metrics in preserving model performance, while the Equivalent Pruning mechanism effectively mitigates performance degradation after fine-tuning. Code: https://github.com/ShaowuChen/Optimal_Brain_Connection

Figures

Figures reproduced from arXiv: 2508.05521 by the authors.

**Figure 2.** Gradients of deep neural networks do not necessar [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** The diagonal blocks of abs(J ⊤J) in the final convolutional layer of ResNet-56 are dominant. Note that here, J ⊤J is block-diagonal, not diagonal. (Best viewed in color and zoom.) Calculating the full J ⊤J is computationally inefficient due to the vast number of parameters (ranging from millions to billions). To address the issue, we assume that only intracomponent parameters (such as those within the same filter … view at source ↗

Figures from the paper (6 more)

**Figure 4.** Figure 4: The J ⊤ mJm of the 16 filters in Conv1 of ResNet-56 shows distinct patterns. Criterion Formula ℓ2 norm ∥wm∥ 2 2 = w⊤ mIwm Taylor P i (wigi) 2 = w⊤ m [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

**Figure 5.** Figure 5: Pruned results on VGG19 for CIFAR-100 without fine-tuning. Vertical lines represent standard deviations. 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Pruned Parameters (M) 0.2 0.4 0.6 0.8 1.0 Train Accuracy L1 FPGM BN Scale Random WHC Taylor Hessian Jacobian (a) ResNet-56 for C…

**Figure 6.** Figure 6: Pruned results of various criteria on ResNet [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

**Figure 7.** Figure 7: Pruned results of various criteria on ViT-B/16 for [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]

**Figure 9.** Figure 9: Ablation on JC’s configurations. Here, “Mean [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]

**Figure 10.** Figure 10: Pruned YOLOv7 without fine-tuning. Fine-tuning % 1 epoch mAP@ 0.5 0.5:0.95 0.5 0.5:0.95 Hessian (Liu et al. 2021) 0.007 0.004 0.026 0.015 Jacobian (ours) 0.193 0.120 0.276 0.174 [PITH_FULL_IMAGE:figures/full_fig_p007_10.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

CoCurve: Cross-Module Co-Pruning Curvature for Training-Free Structured LLM Pruning
cs.LG 2026-07 conditional novelty 6.0 of 10

A training-free joint attention+FFN pruning method whose Fisher-derived interaction matrix, recovered from M single-unit ablations as a Gram product, beats node-only saliency baselines on 3B–24B models.
Adaptive Parallel Downloader for Large Genomic Datasets
cs.DC 2025-08 unverdicted novelty 5.0 of 10

FastBioDL uses utility-based gradient descent to adapt HTTP/FTP download concurrency in real time, reporting up to 4x faster retrieval of large genomic datasets than static-concurrency tools.

Reference graph

Works this paper leans on

7 extracted references · 6 canonical work pages · cited by 2 Pith papers

[4]

Efficient Model Compression Techniques with FishLeg

Learning Efficient Convolutional Networks through Network Slimming. In Proceedings of the IEEE Interna- tional Conference on Computer Vision (ICCV), 2755–2763. Liu, Z.; Mu, H.; Zhang, X.; Guo, Z.; Yang, X.; Cheng, K.; and Sun, J. 2019. MetaPruning: Meta Learning for Auto- matic Neural Network Channel Pruning. In Proceedings of the IEEE International Confe...

work page Pith review arXiv 2019
[6]

In International Conference on Learn- ing Representations (ICLR)

Pruning Convolutional Neural Networks for Resource Efficient Inference. In International Conference on Learn- ing Representations (ICLR). Nonnenmacher, M.; Pfeil, T.; Steinwart, I.; and Reeb, D

work page
[2017]

In Interna- tional Conference on Learning Representations (ICLR)

Pruning Filters for Efficient ConvNets. In Interna- tional Conference on Learning Representations (ICLR). Liao, Z.; Hezbri, N.; Qu ´etu, V .; Nguyen, V .-T.; and Tartaglione, E. 2025. Till the Layers Collapse: Compressing a Deep Neural Network through the Lenses of Batch Nor- malization Layers. In Proceedings of the AAAI Conference on Artificial Intellige...

work page 2025
[2019]

In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 11264–11272

Importance Estimation for Neural Network Pruning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 11264–11272. Molchanov, P.; Tyree, S.; Karras, T.; Aila, T.; and Kautz, J

work page
[2020]

AutoSculpt: A Pattern-based Model Auto-pruning Framework Using Reinforcement Learning and Graph Learning

Asymptotic Soft Filter Pruning for Deep Convolu- tional Neural Networks. IEEE Trans. Cybern., 3594–3604. He, Y .; Liu, P.; Wang, Z.; Hu, Z.; and Yang, Y . 2019. Filter Pruning via Geometric Median for Deep Convolutional Neu- ral Networks Acceleration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 4340–4349. H...

work page Pith review arXiv 2019
[2022]

Conv-BN-CCC-New BN-ReLU-DDD-Conv/Classifier

SOSP: Efficiently capturing global correlations by second-order structured pruning. International Conference on Learning Representations (ICLR). Singh, P.; Verma, V . K.; Rai, P.; and Namboodiri, V . P. 2020. Leveraging Filter Correlations for Deep Model Compres- sion. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WA...

arXiv 2020
[2025]

In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), 3666–3675

Text-guided Sparse V oxel Pruning for Efficient 3D Visual Grounding. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), 3666–3675. Han, S.; Pool, J.; Tran, J.; and Dally, W. J. 2015. Learn- ing Both Weights and Connections for Efficient Neural Net- work. In Advances in Neural Information Processing Sys- tems (Ne...

work page 2015

Pith tools

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

[1] [4]

Efficient Model Compression Techniques with FishLeg

Learning Efficient Convolutional Networks through Network Slimming. In Proceedings of the IEEE Interna- tional Conference on Computer Vision (ICCV), 2755–2763. Liu, Z.; Mu, H.; Zhang, X.; Guo, Z.; Yang, X.; Cheng, K.; and Sun, J. 2019. MetaPruning: Meta Learning for Auto- matic Neural Network Channel Pruning. In Proceedings of the IEEE International Confe...

work page Pith review arXiv 2019

[2] [6]

In International Conference on Learn- ing Representations (ICLR)

Pruning Convolutional Neural Networks for Resource Efficient Inference. In International Conference on Learn- ing Representations (ICLR). Nonnenmacher, M.; Pfeil, T.; Steinwart, I.; and Reeb, D

work page

[3] [2017]

In Interna- tional Conference on Learning Representations (ICLR)

Pruning Filters for Efficient ConvNets. In Interna- tional Conference on Learning Representations (ICLR). Liao, Z.; Hezbri, N.; Qu ´etu, V .; Nguyen, V .-T.; and Tartaglione, E. 2025. Till the Layers Collapse: Compressing a Deep Neural Network through the Lenses of Batch Nor- malization Layers. In Proceedings of the AAAI Conference on Artificial Intellige...

work page 2025

[4] [2019]

In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 11264–11272

Importance Estimation for Neural Network Pruning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 11264–11272. Molchanov, P.; Tyree, S.; Karras, T.; Aila, T.; and Kautz, J

work page

[5] [2020]

AutoSculpt: A Pattern-based Model Auto-pruning Framework Using Reinforcement Learning and Graph Learning

Asymptotic Soft Filter Pruning for Deep Convolu- tional Neural Networks. IEEE Trans. Cybern., 3594–3604. He, Y .; Liu, P.; Wang, Z.; Hu, Z.; and Yang, Y . 2019. Filter Pruning via Geometric Median for Deep Convolutional Neu- ral Networks Acceleration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 4340–4349. H...

work page Pith review arXiv 2019

[6] [2022]

Conv-BN-CCC-New BN-ReLU-DDD-Conv/Classifier

SOSP: Efficiently capturing global correlations by second-order structured pruning. International Conference on Learning Representations (ICLR). Singh, P.; Verma, V . K.; Rai, P.; and Namboodiri, V . P. 2020. Leveraging Filter Correlations for Deep Model Compres- sion. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WA...

arXiv 2020

[7] [2025]

In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), 3666–3675

Text-guided Sparse V oxel Pruning for Efficient 3D Visual Grounding. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), 3666–3675. Han, S.; Pool, J.; Tran, J.; and Dally, W. J. 2015. Learn- ing Both Weights and Connections for Efficient Neural Net- work. In Advances in Neural Information Processing Sys- tems (Ne...

work page 2015