WorksComputer Science
Structured Channel Network Pruning with Bi-Sparse Matrix Factorization
Abstract
Structured channel pruning for convolutional neural networks (CNNs) plays an important role in pattern recognition and deep learning. Nevertheless, existing methods usually calculate channel importance scores based on the magnitudes of individual filters, making it difficult to distinguish shared redundancy from individual channel contributions. To tackle this limitation, we propose a novel structured channel pruning method termed bi-sparse matrix factorization (BSMF). Specifically, BSMF characterizes the convolutional weight matrix as output-channel coefficients over a shared basis with row-wise -norm and column-wise -norm regularization terms, respectively. Note that , which, to our best knowledge, has not been explored in the literature. To solve the resulting non-convex optimization problem, we develop an efficient update scheme built upon the alternating direction method of multipliers (ADMM), in which all subproblems admit closed-form solutions. After that, a compact CNN is obtained by reconstructing the convolutional filters from the learned factors and discarding the redundant channels. Extensive experiments on the CIFAR-10 and ImageNet-1K datasets demonstrate that our proposed BSMF achieves significant network compression while maintaining high reconstruction accuracy. The source code is publicly available at https://github.com/xianchaoxiu/BSMF
Email addresses: [email protected] (Jingjing Liu), [email protected] (Ziyang Wang), [email protected] (Xianchao Xiu), [email protected] (Wanquan Liu)
Keywords: Network compression, structured pruning, convolutional neural networks, matrix factorization, sparse optimization
Introduction
Convolutional neural networks (CNNs) have achieved considerable success in computer vision [1]. However, when deployed on mobile devices, deeper and wider networks must operate within limited memory and computation resources, which limits their practical implementation [2]. In this regard, model compression techniques have been developed to reduce network complexity while preserving predictive accuracy. Interested readers are referred to recent surveys [3, 4].
Generally speaking, existing model compression methods are grouped into four categories: pruning, quantization, knowledge distillation, and low-rank decomposition. Among them, pruning eliminates redundant weights or structured units [5], quantization reduces the precision of weights or activations [6], knowledge distillation trains a compact student model under the supervision of a teacher model [7], and low-rank decomposition factorizes dense weight matrices or tensors into smaller components [8].
Compared with the other three categories, pruning possesses the advantage of achieving flexible trade-offs between model performance and computational complexity. As discussed in [9], network pruning methods fall into unstructured and structured pruning. Technically, unstructured pruning produces irregular sparsity and usually demands dedicated software or hardware support for speedup. By contrast, structured pruning discards complete convolutional filters, yielding a compact network compatible with standard dense convolution implementations [10]. Consequently, structured pruning has garnered significant attention from both academia and industry [11].
Within structured pruning, filters are commonly selected using cross-layer importance evaluation [12], global channel attention [13], or gradient information [14]. Other paradigms include channel masks [15] and dynamic strategies realized by input-dependent gates [16]. Nevertheless, scores based on individual filter responses may fail to characterize shared components and channel-wise contributions. Recent efforts have sought to capture interdependencies among filters using independence criteria or tensor decomposition techniques [17, 18]. For example, efficient tensor decomposition-based filter pruning represented individual filters using higher-order singular value decomposition and identified similar filters from distances between their factors [18]. To exploit filter similarity during optimization, co-synthesizing multiple similar filters (CSMF) grouped similar filters and synchronized their updates using fused gradients and individual weight increments, consolidating their information into representative filters before removal [10]. More recently, dynamic filter pruning via unified importance and redundancy combined filter magnitudes with inter-filter correlations to consider filter importance and redundancy [19]. Although these methods alleviate inter-filter redundancy, they do not learn output-channel coefficients over a shared basis. Therefore, by aligning the coefficients in the decomposed representation with the output channels, one can build an explicit connection between shared structural redundancies and channel selection.
A complementary strategy is to use sparse regularization terms to induce removable channel structures. Specifically, network slimming using -norm () regularization term approximated -norm sparsity [5], while channel pruning for convolutional neural networks using -norm constraints directly constrained nonzero channel indicators [20]. Although these objectives facilitate channel removal, they do not explicitly model the shared structure among filters. This motivates a combined treatment of shared redundancy and channel sparsity. In practice, channel selection must satisfy various constraints, such as layer-width requirements, residual connections, and global resource budgets [21, 22]. At the deployment stage, some methods retain the decomposed factors as separate operators during inference, which can introduce additional computational overhead [23]. Therefore, a practical channel pruning framework should connect shared redundancy modeling with channel selection while preserving standard dense convolutions for deployment, without introducing additional factorized operators during inference.
Motivated by the above observations, we propose bi-sparse matrix factorization (BSMF) for structured channel pruning. This method decomposes each convolutional weight matrix into channel coefficients and a shared basis, where each output channel corresponds to a specific coefficient vector. In addition, we impose structured sparsity on the rows of the coefficient factor and the columns of the basis factor. Here, the row-wise norms of the coefficient are used as a scoring criterion to select channels, while adhering to constraints on layer width, connectivity, and computational resources. Before removing the channels, the factors are recombined to reconstruct the convolutional filters, ultimately yielding a compact network composed of standard dense convolutions. Figure 1 illustrates the differences between our proposed BSMF and existing CSMF [10]. It can be seen that CSMF coordinates the update process of filters by combining a common fusion gradient with individual weight increments, while our BSMF constructs channel scores by utilizing shared convolutional filter information, and each output filter is represented by a row of coefficients over the shared basis. Furthermore, the norm of this row determines the priority for retaining the corresponding filter, which provides greater flexibility.

Figure 1. Comparison between our BSMF and CSMF [10], where BSMF scores channels using the row-wise norms of , while CSMF co-synthesizes filters using fused gradients and individual weight increments.
The main contributions of this paper can be summarized as follows.
We propose a novel matrix factorization model by imposing non-convex structured regularization on both the output-channel coefficients and the shared basis.
By leveraging the row-wise norms of the learned coefficients, we derive channel importance scores that account for filter relationships, thereby providing guidance for structured channel pruning.
We develop an efficient optimization algorithm based on the alternating minimization framework, where both variables can be solved using ADMM. Although establishing strict theoretical convergence is challenging, we empirically verify the convergence numerically.
We demonstrate the superiority of the proposed method through sufficient experiments. Moreover, we visualize the learned importance scores using channel-selection heatmaps and quantify the selection differences by comparing the sets of retained channels via the Jaccard coefficient.
The remainder of this paper is organized as follows. Section 2 reviews the related work. Section 3 presents our mathematical formulation, optimization algorithm, and pruning procedure. Section 4 reports the experimental setup, numerical results, ablation studies, and further discussions. Section 5 concludes the paper.
Related Work
This section reviews recent work on channel pruning, structured pruning, and sparse regularization.
Channel Pruning
Channel scoring must distinguish information content from inter-channel redundancy. In previous studies, filter pruning using high-rank feature map ranks feature responses but may overlook cross-channel duplication [24]. Later, channel independence-based pruning [25] and automatic network pruning via Hilbert-Schmidt independence criterion lasso under information bottleneck principle [26] capture channel dependence or information preservation. However, these methods rely on activation statistics derived from sampling when assigning scores. Although the above methods reduce redundancy, they do not learn sparse channel coefficients within a shared-basis framework. Beyond similarity-based pruning, more recent methods including singular value-driven automated filter pruning [27], adding before pruning (ABP) [28], and iterative pruning with prompt knowledge distillation (IPPKD) [29] preserve information through spectral structure, attention-guided fusion, or distillation, respectively. By contrast, our BSMF uses the row-wise norms of the shared-basis coefficients to score channels, aligning scoring with sparse optimization yet without activation statistics.
Structured Pruning
Structured pruning removes complete filters, and existing methods mainly rely on masks, filter similarity, or factorized representations. Although generative adversarial learning (GAL) [30] and leanflex grouped filter pruning (LeanFlex-GKP) [31] learn structural masks or group configurations, they do not explicitly represent shared filter components. Besides, tensor decomposition-based filter pruning [18] and CSMF [10] characterize inter-filter similarity through decomposition-factor distances and coordinated fusion, respectively. However, these methods do not learn channel-aligned sparse coefficients over a shared basis [23]. Our BSMF addresses these limitations by learning a shared representation for channel selection and recombining its factors before removal, thereby preserving standard dense convolutions at deployment.
Sparse Regularization
Sparse regularization methods also differ in the granularity of the structures they sparsify. Beyond explicit channel regularizations, feature flow regularization promotes structured sparsity through feature propagation [32]. More recently, double sparse factorization learns two element-wise sparse factor matrices [33]. Nevertheless, feature-level or element-wise sparsity does not necessarily eliminate an entire output-channel group. In contrast, our BSMF couples weight reconstruction with -norm regularization on coefficient rows and -norm regularization on basis columns. The row regularization suppresses complete output-channel groups, whereas the basis-column regularization encourages a compact shared representation.
Methodology
This section first introduces the proposed formulation and optimization algorithm, followed by the pruning strategy.
Our Formulation
For a convolutional layer, unfold its weights into with and , where and denote the number of output and input channels, respectively, and represents the width of the convolution filter. For the sake of convenience, we omit the layer index and approximate the weight matrix as
where and , with integer rank . In fact, each row of corresponds to an output channel, while the rows of form a shared basis and its columns correspond to the input-filter coordinates.
To enhance pruning capability, we consider the following model
where is the Frobenius norm, are parameters to weight the two sparsity-inducing regularization terms. We would like to point out that
where denotes the -row of and denotes the -th column of . This means that are row-wise and column-wise structured sparsity, respectively. In addition, , which includes in [34] and the special case in [35]. Specifically, when , they are defined as the counts of nonzero rows and nonzero columns, respectively.
Optimization Algorithm
Problem (3) involves two variables, making direct optimization quite difficult. Inspired by [33], we adopt an alternating minimization scheme by computing
In what follows, we describe how to solve the subproblems using the alternating direction method of multipliers (ADMM).
A-subproblem
With fixed, A-subproblem can be reformulated as
The corresponding augmented Lagrangian is
where denotes the inner product and is the tuning parameter.
Update according to
Setting the gradient with respect to to zero yields
where denotes the transpose of . Therefore, the closed-form solution of problem (7) is
where is the identity matrix.
Update by
where . Thus, the closed-form is
Specifically, for the -th row, it becomes
and prox is defined in the following lemma.
Lemma 3.1. For and , consider the proximal operator
where
More details can be found in [36].
Update via
B-subproblem
With the updated fixed, -subproblem can be written as
and the augmented Lagrangian is
Similar to the above updates, the following iterative forms can be obtained.
Update by
where where denotes the transpose of .
Update by
where and its column is defined as
Algorithm 1 Alternating Minimization for Solving Problem (3)
Input: Weight , rank , iteration , parameters , , , , .
Initialize: Compute , let , , , .
while do
A-subproblem
Update using Eq. (9).
Update using Eq. (11).
Update using Eq. (15).
end
B-subproblem
Update using Eq. (18).
Update using Eq. (19).
Update using Eq. (21).
end
.
end
Output: , and compute using Eq. (22).
Update by
Overall Scheme
After outer iterations, the weight matrix is reconstructed as
Finally, Algorithm 1 summarizes the initialization and alternating updates.
Pruning Strategy
This part describes how to use the factor matrix to rank output channels, construct masks, and then convert the masked model into a compact dense network.
Channel Importance
Let denote the number of output channels in the -th layer. Its importance score is defined as
A smaller score indicates a weaker coefficient group and a lower retention priority. Since the raw score scales may vary across different layers, the scores are normalized independently for each layer. Let and be the 5th and 95th percentiles of the scores in the -th layer, respectively. When , define
where and . This robust scaling limits the effect of extreme scores while placing layers on a common range. If , all normalized scores for that layer are set to one, thereby avoiding division by an extremely small value.
Mask Selection
Let be the set of layers eligible for output channel pruning, and let be the total number of channels across these layers. For a given nominal pruning ratio , the normalized scores of all eligible channels are sorted in ascending order as . In addition, the initial threshold is defined as the following empirical lower-tail order statistic
The initial mask is defined as
where 1 indicates that the channel is retained, and 0 indicates that it is marked for removal. When all channel scores are distinct, this threshold results in the removal of the channels with the lowest scores, and if any channels have scores equal to the threshold, they are also removed. Thus, determines the nominal lower-tail proportion. Subsequently, the threshold mask must be refined to satisfy constraints regarding layer width and connectivity. Let and be the allowed minimum and maximum retained widths, where . If a layer retains fewer than channels, its highest-scoring discarded channels are restored until the lower bound is met. If it retains more than , its lowest-scoring retained channels are removed until the upper bound is met. Let denote the resulting mask and be the corresponding retained index set. At the adjustment boundaries, if multiple indices share the same score, it is possible to derive several valid masks. Furthermore, if removing certain channels would violate the width constraints imposed by residual addition operations, those channels are protected and retained.
Network Construction
For each layer, the dense convolutional filter is reconstructed using Eq. (22), and the output filters indexed by are retained. Simultaneously, the associated normalization parameters and input-channel slices are removed from the subsequent connected layers. After that, the batch-normalization statistics are recalibrated, and the compact model is fine-tuned to recover its accuracy.
Experiments
To validate the superiority of our proposed BSMF, some state-of-the-art pruning methods are selected, including HRank2, CHIP3, APIB4, CORING5, NORTON6, CSMF7, DFPV8, GAL9, LeanFlex-GKP10, 11 12, ABP, and IPPKD13.
Subsection 4.1 describes the datasets, models, and training settings. Subsection 4.2 evaluates the proposed BSMF on the CIFAR-10 and ImageNet-1K datasets in terms of accuracy, floating-point operations (FLOPs), and the number of parameters. Subsection 4.3 examines the roles of individual components, including matrix factorization, bi-sparsity optimization, and channel scoring. Subsection 4.4 provides further analysis of the convergence, resulting channel selections, sensitivity to the regularization order, and distribution of effective weights.
Experimental Setup
Datasets and Models
Experiments are conducted on the CIFAR-1014 dataset using ResNet-56, ResNet-110, and VGG-16 models, while the ResNet-50 model is used for the ImageNet-1K15 dataset. Here, the CIFAR-10 dataset consists of 50,000 training images and 10,000 testing images across 10 classes, with an image size of pixels. The ImageNet-1K dataset contains approximately 1.28 million training images and 50,000 validation images across 1,000 classes.
Training Settings
During the fine-tuning stage, stochastic gradient descent (SGD) is employed. For the CIFAR-10 dataset, the batch size is set to 128 for the ResNet-56, ResNet-110, and VGG-16 models. The learning rates for ResNet-56 at 40% FLOPs reduction, ResNet-110 at approximately 50% FLOPs reduction, and VGG-16 are 0.009, 0.01, and 0.015, respectively. The corresponding momentum values are 0.95, 0.95, and 0.97, with weight decay for all three. For the ImageNet-1K dataset, ResNet-50 is fine-tuned using Nesterov SGD with a batch size of 256, a learning rate of 0.025, a momentum of 0.9, and a weight decay coefficient of . Furthermore, normalization parameters and bias terms are excluded from weight decay. For ResNet-56 and ResNet-110, all comparison methods are reimplemented with their recommended hyperparameters, using the same pretrained checkpoint, pruning targets, and fine-tuning protocol within each architecture. For VGG-16 and ResNet-50, the comparison results are taken directly from the corresponding papers, together with their original training recipes and dense baselines.
Evaluation Metrics
To evaluate the performance of the comparative methods, Top-1 and Top-5 accuracy are used, representing the proportion of cases where the ground-truth label appears within the top 1 and top 5 categories, respectively, based on the predicted scores [37]. A higher value indicates better performance. Additionally, FLOPs and the number of parameters (Params) [24] are considered, where lower values are preferred when accuracy is comparable. For any of the aforementioned resource metrics, the pruning ratio (PR) is calculated as
where and are the corresponding counts for the dense baseline and pruned model. A higher PR value indicates a greater relative reduction.
Comparison Results
Table 1 shows that our BSMF achieves the highest accuracy across the evaluated pruning settings. Its gain over the strongest competing method increases from 0.26% and 0.01% under mild pruning to 0.63% and 0.47% under tight pruning on ResNet-56 and ResNet-110, respectively. Its shared-representation-based channel scoring captures inter-filter dependencies, helping BSMF retain informative and complementary channels under tight pruning budgets, thereby retaining a more informative channel subset under the same pruning budget. Under mild pruning, ranking differences have limited impact because many redundant channels remain. Under tight pruning, however, an incorrect ranking causes a larger loss of feature coverage, making channel-selection quality decisive. The ablation results in Subsection 4.3 support this interpretation. Overall, our BSMF achieves a more favorable accuracy-efficiency trade-off under comparable FLOPs and parameter constraints.
| Model | Method | Top-1 Accuracy | FLOPs (PR) | Params (PR) |
| ResNet-56 | Baseline | 93.26% | 127.09M (0.0%) | 0.85M (0.0%) |
| [5] | 93.87% | 75.62M (40%) | 0.56M (34%) | |
| CORING [18] | 93.79% | 76.12M (40%) | 0.47M (45%) | |
| APIB [26] | 93.92% | 76.35M (40%) | 0.53M (37%) | |
| BSMF (Ours) | 94.18% | 75.92M (40%) | 0.44M (48%) | |
| HRank [24] | 92.49% | 62.72M (50%) | 0.49M (42%) | |
| NORTON [23] | 92.41% | 63.33M (50%) | 0.38M (55%) | |
| Pruner [20] | 93.03% | 64.04M (50%) | 0.36M (58%) | |
| BSMF (Ours) | 93.84% | 62.23M (50%) | 0.34M (60%) | |
| ResNet-110 | Baseline | 93.50% | 257.09M (0.0%) | 1.73M (0.0%) |
| [5] | 94.41% | 126.70M (50%) | 1.11M (36%) | |
| APIB [26] | 94.45% | 127.29M (50%) | 0.88M (49%) | |
| BSMF (Ours) | 94.46% | 125.00M (51%) | 0.89M (48%) | |
| Pruner [20] | 93.26% | 114.92M (55%) | 0.69M (60%) | |
| HRank [24] | 92.96% | 105.63M (58%) | 0.72M (60%) | |
| NORTON [23] | 93.58% | 102.84M (60%) | 0.65M (62%) | |
| BSMF (Ours) | 94.05% | 102.04M (60%) | 0.69M (60%) |
Table 1. Pruning results for ResNet-56 and ResNet-110 on the CIFAR-10 dataset, where the best and second-best pruned values are shown in bold and underlined, respectively.
Table 2 reports the pruning results of VGG-16 on the CIFAR-10 dataset. Under the high-compression setting, our BSMF achieves an accuracy of 94.48%, the highest among the compared methods, while substantially reducing both computation and model size. This result shows that BSMF preserves classification performance even when pruning is aggressive. Together with the results on ResNet-56 and ResNet-110, it demonstrates that our BSMF is effective on the CIFAR-10 dataset across both sequential architectures such as VGG-16 and residual architectures, rather than being limited to a particular network design.
| Model | Method | Top-1 Accuracy | FLOPs (PR) | Params (PR) |
| VGG-16 | Baseline | 93.96% | 313.73M (0%) | 14.98M (0%) |
| LeanFlex-GKP [31] | 94.15% | 178.17M (43%) | 8.29M (44%) | |
| GAL-0.1 [30] | 93.42% | 171.89M (45%) | 2.67M (82%) | |
| HRank [24] | 93.43% | 145.61M (54%) | 2.51M (83%) | |
| DFPV [19] | 93.76% | 139.30M (56%) | 3.25M (78%) | |
| IPPKD [29] | 94.24% | 139.36M (56%) | 6.89M (54%) | |
| CHIP [25] | 93.86% | 131.17M (58%) | 2.76M (82%) | |
| CORING [18] | 94.42% | 131.44M (58%) | 2.76M (82%) | |
| APIB [26] | 94.08% | 125.48M (60%) | 3.60M (76%) | |
| [5] | 94.36% | 123.60M (61%) | 2.67M (82%) | |
| BSMF (Ours) | 94.48% | 122.98M (61%) | 2.66M (82%) |
Table 2. Pruning results for VGG-16 on the CIFAR-10 dataset, where the best and second-best pruned values are shown in bold and underlined, respectively.
Furthermore, Table 3 presents the pruning results for ResNet-50 on the ImageNet-1K dataset. Our BSMF achieves the highest reported Top-1 accuracy of 76.27%, exceeding CHIP by 0.12% while requiring fewer FLOPs (2.07B versus 2.11B) and parameters (13.03M versus 14.23M). This performance is consistent with the coupling of reconstruction and sparse channel selection in BSMF. The reconstruction term penalizes deviations from the original weights, while row-wise regularization suppresses weak channel coefficients in the shared basis. Ranking channels by these regularized coefficient norms carries the learned suppression directly into pruning decisions, aligning channel selection with sparse optimization. Together with the CIFAR-10 results, this finding supports the effectiveness of BSMF across the evaluated architectures and its applicability to large-scale image classification.
Table 3: Pruning results for ResNet-50 on the ImageNet-1K dataset, where the best and second-best pruned values are shown in bold and underlined, respectively.
| Model | Method | Top-1 Accuracy | Top-5 Accuracy | FLOPs (PR) | Params (PR) |
| ResNet-50 | Baseline [38] | 76.15% | 92.87% | 4.12B (0.0%) | 25.50M (0.0%) |
| LeanFlex-GKP [31] | 75.62% | N/A | 2.76B (33%) | 17.80M (30%) | |
| GAL-0.5 [30] | 71.95% | 90.94% | 2.33B (43%) | 21.20M (17%) | |
| ABP [28] | 74.80% | 92.36% | 2.36B (43%) | 13.01M (49%) | |
| HRank-1 [24] | 74.98% | 92.33% | 2.30B (44%) | 16.15M (37%) | |
| Pruner [20] | 75.18% | 92.39% | 2.27B (45%) | 13.39M (48%) | |
| [5] | 76.06% | _92.91%_ | 2.22B (46%) | 17.14M (33%) | |
| CSMF [10] | 75.81% | N/A | 2.20B (47%) | 14.92M (42%) | |
| CHIP [25] | _76.15%_ | _92.91%_ | 2.11B (49%) | 14.23M (44%) | |
| DFPV [19] | 75.89% | 93.04% | 1.99B (52%) | 14.05M (45%) | |
| BSMF (Ours) | 76.27% | 92.89% | 2.07B (50%) | _13.03M (49%)_ |
Table 3. Pruning results for ResNet-50 on the ImageNet-1K dataset, where the best and second-best pruned values are shown in bold and underlined, respectively.
Overall, the results on the CIFAR-10 and ImageNet-1K datasets demonstrate that our BSMF achieves a favorable accuracy-efficiency trade-off across both sequential and residual architectures. The larger accuracy gains under more aggressive pruning budgets on ResNet-56 and ResNet-110 further show its effectiveness when the retained model capacity is limited.
Ablation Study
The ablation study compares factorization, bi-sparse optimization, and channel-scoring variants on the CIFAR-10 dataset, then examines their scores and retained-channel sets to explain the resulting selections.
Component Contributions
Table 4 reports the ResNet-56 ablation results on the CIFAR-10 dataset, where accuracy is reported as the mean and standard deviation over three random seeds. It can be concluded that
Table 4: Component ablation results for ResNet-56 on the CIFAR-10 dataset, where the best and second-best pruned values are shown in bold and underlined, respectively.
| Method | Factor. | Sparse opt. | Channel score | Top-1 Accuracy (%) | FLOPs Red. | Params |
| Baseline | No | No | None | 93.26 | 0% | 0.85M |
| BSMF without pruning | Yes | Yes | None | 93.09 ± 0.02 | 0% | 0.85M |
| pruning | No | No | filter -norm | 93.65 ± 0.23 | 40% | 0.55M |
| BSMF + filter | Yes | Yes | filter -norm | 93.74 ± 0.10 | 40% | 0.52M |
| BSMF + filter | Yes | Yes | filter -norm | _93.91 ± 0.09_ | 40% | _0.51M_ |
| Factorization-only pruning | Yes | No | A-row norm | 93.83 ± 0.18 | 40% | 0.44M |
| BSMF (Ours) | Yes | Yes | A-row norm | 93.99 ± 0.07 | 40% | 0.44M |
Table 4. Component ablation results for ResNet-56 on the CIFAR-10 dataset, where the best and second-best pruned values are shown in bold and underlined, respectively.
Without pruning, BSMF achieves 93.09% accuracy versus 93.26% for the dense baseline, with unchanged FLOPs and 0.85M parameters. The 0.17-percentage-point decrease shows that factorization, bi-sparse optimization, and filter recovery preserve near-baseline performance before channel removal.
With identical -row scoring, 40% FLOPs reduction, and 0.44M parameters, adding bi-sparse optimization to factorization-only pruning raises mean accuracy from 93.83% to 93.99% and reduces the observed standard deviation from 0.18 to 0.07, supporting its effectiveness beyond factorization alone.
At 40% FLOPs reduction, adding factorization and bi-sparse optimization to direct pruning raises mean accuracy from 93.65% to 93.74% and reduces parameters from 0.55M to 0.52M. This modest benefit with recovered-filter scoring supports the pipeline’s effectiveness and provides a reference for evaluating -row scoring.
Among the three BSMF scoring variants, filter- improves accuracy over filter- from 93.74% to 93.91%, while -row scoring achieves the highest accuracy (93.99%) with the fewest parameters (0.44M). By scoring the regularized coefficient groups rather than recovered filters, -row scoring directly connects sparse optimization to channel selection, yielding the best pruning performance among the evaluated configurations.
Overall, BSMF maintains near-baseline accuracy before pruning, benefits from bi-sparse optimization, and improves conventional norm-based pruning. Notably, with -row scoring, it achieves the highest accuracy with the fewest parameters among the evaluated configurations.
Effect of Factorized Channel Scoring
Figure 2 compares baseline filter -norms (a), optimized -row scores (b), and the final retention mask (c) for ResNet-56. The scores in (b), defined in Eq. (23), measure each channel’s coefficient strength in a shared basis. Compared with (a), they show clearer contrast among channels, particularly in the deeper stages, consistent with row-wise regularization suppressing weak coefficient groups. The mask in (c) is constructed from these scores under the network’s structural constraints, linking the regularized factorized representation to channel selection.
Figure 2. Channel-selection heatmaps for ResNet-56, where (a)-(c) show baseline filter -norms, BSMF A-row scores, and retention masks, respectively.
Comparison of Retained-Channel Sets
To examine how its channel selections differ from those of existing criteria, we compare the retained-channel sets at the same layer using the Jaccard similarity. For layer , it is defined as
where and are the channel sets retained by methods and , respectively. Figure 3 shows pairwise similarities for representative layers of ResNet-56 and ResNet-19
Figure 3. Pairwise Jaccard similarity of retained-channel sets on the CIFAR-10 dataset, where the top and bottom rows show ResNet-56 and ResNet-110, respectively.
The mean-similarity plots indicate that BSMF’s selections are generally less similar to those of the other criteria in the middle and deeper layers, while the trend in shallow layers varies between the two networks. This suggests that BSMF can retain channel sets that differ from those selected by conventional criteria, consistent with its factorized channel-scoring mechanism. However, the reasons for the layer-dependent differences in similarity remain unclear and warrant further investigation.
The ablations show higher mean accuracy with bi-sparse optimization than with factorization alone at the tested budget, while the retained-set comparisons reveal selection differences from the evaluated criteria. BSMF’s key feature is the use of regularized, channel-aligned coefficient groups in a learned shared basis for importance scoring, directly linking sparse optimization to channel selection.
Discussion
This section provides a convergence analysis, examines how regularization orders and architectural constraints affect pruning, and uses effective weight distributions to relate factor sparsity to channel selection.
Convergence Analysis
Figure 4(a)-(Figure 4d) shows the objective values of problem (3) during Algorithm 1 for VGG-16, ResNet-56, and ResNet-110 on the CIFAR-10 dataset, and ResNet-50 on the ImageNet-1K dataset, respectively. Across all four settings, the objective decreases rapidly in the first few shrink iterations and then approaches a stable level, with only small subsequent changes. BSMF couples the row- and column-group penalties in Eq. (3) to regularize channel coefficients and the shared basis, respectively. The observed stabilization provides numerical support for optimizing this coupled nonconvex formulation through alternating ADMM.

Figure 4. Convergence curves of BSMF on different models and datasets, where (a)-(d) show the objective function values.
Sensitivity to Regularization Orders
We evaluate nine pairs for the regularization orders in Eq. (3) on ResNet-56 and ResNet-110 using the CIFAR-10 dataset. Figure 5 reports the mean Top-1 accuracy over three random seeds. The results suggest selecting and jointly. For ResNet-110, generally performs well, suggesting half-order regularization on the shared-basis side while considering several candidate orders on the coefficient side. Here, and govern row sparsity in channel coefficients and column sparsity in the shared basis, respectively. Overall, the mean accuracy ranges span 0.28 and 0.24 percentage points for ResNet-56 and ResNet-110, respectively, suggesting relative insensitivity to these orders within the tested range. However, the limited grid and small differences between configurations require further evaluation with more seeds and combined consideration of regularization weights, factorization rank, and fine-tuning settings to determine preferred choices. Their applicability to other models also requires validation.
Figure 5. Sensitivity to and on the CIFAR-10 dataset, where (a,b) show mean best-checkpoint Top-1 accuracy over three seeds for ResNet-56 and ResNet-110, respectively.
Channel Allocation and Deployment
Figure 6 illustrates channel protection by layer function and pruning sensitivity [22]. VGG-16 (a) compresses earlier layers while preserving its final output width for the classifier, consistent with ThiNet-Tiny’s milder final-layer pruning [39]. ResNet-56 and 110 (b,c) retain block output widths for shortcut compatibility and protect the stem and Stage 1 against early information loss, with stronger protection at the entrances of Stages 2 and 3 for cross-scale transfer and at the end of Stage 3 for classifier inputs. ResNet-50 (d) prunes bottleneck interiors while retaining block outputs and projection shortcuts, as in ThiNet [39]. These constraints preserve feature transfer, residual addition, and classification while concentrating compression on internal computation. Within these constraints, BSMF selects channels using coefficient-row scores learned through bi-sparse shared filter modeling, then recombines the factors and physically removes channels to obtain a compact network with standard dense convolutions.
Figure 6. Channel counts before and after pruning, where counts are shown per VGG-16 layer and summed over the main-branch convolutions of each ResNet block.
Effective Weight Distributions
Figure 7 compares original weights (a)-(d) with effective weights (e)-(h), defined for visualization as , where contains dense filters after factor reconstruction and recovery, and uses the normalized -row score in Eq. (24). Across all four architectures, effective weights concentrate more strongly near zero, as illustrated in (a,e). Since each row represents an output channel, score-based attenuation visualizes the learned coefficient-row sparsity. Under structural constraints, BSMF converts these scores into pruning masks, translating the sparse structure learned through bi-sparse shared filter modeling into a channel-pruning strategy.

Figure 7. Weight distributions, where (a)-(d) show original weights and (e)-(h) show effective weights constructed using normalized A-row scores.
Conclusion
This paper proposes a novel structured channel pruning method called BSMF. By expressing convolutional filters as channel-aligned coefficients over a shared basis and applying structured sparsity regularization terms to both factors, our BSMF assesses channel importance within a representation that captures relationships among filters. In algorithms, an efficient alternating minimization strategy is designed based on the ADMM framework. Experiments on the CIFAR-10 and ImageNet-1K datasets demonstrate competitive classification accuracy with reduced FLOPs and parameter counts. These results confirm the significant potential of our proposed pruning model.
Of course, there remain some limitations. On one hand, the compression stage requires a significant number of iterative computations, thus a faster and more efficient algorithm should be developed. On the other hand, this work focuses exclusively on CNN-based image models and extending it to large language models would also be a meaningful direction.
Acknowledgments
This work was supported in part by the National Natural Science Foundation of China under Grants 62204044 and 12371306, the State Key Laboratory of Integrated Chips and Systems under Grant SKLICS-K202302, and the Shenzhen Nature Science Research Foundation Key Project under Grant ZDYJ20251211121303005.
References
- [1]X. Zhao, L. Wang, Y. Zhang, X. Han, M. Deveci, M. Parmar, A review of convolutional neural networks in computer vision, Artificial Intelligence Review 57 (2024) 99.DOI
- [2]Z. Li, F. Liu, W. Yang, S. Peng, J. Zhou, A survey of convolutional neural networks: Analysis, applications, and prospects, IEEE Transactions on Neural Networks and Learning Systems 33 (2022) 6999–7019.DOI
- [3]D. Liu, Y. Zhu, Z. Liu, Y. Liu, C. Han, J. Tian, R. Li, W. Yi, A survey of model compression techniques: Past, present, and future, Frontiers in Robotics and AI 12 (2025) 1518965.DOI
- [4]F. M. A. Khan, M. Hallaq, H. Abou-Zeid, O. Erak, O. Waqar, S. A. Hassan, O. Alhussein, E. Hossain, Model compression for sustainable AI in xG wireless networks: Recent advances, challenges, and future directions, IEEE Communications Surveys & Tutorials (2026).
- [5]F. Ji, X. Chen, R. Chu, B. Liu, Network slimming using Lₚ (p < 1) regularization, Pattern Recognition 167 (2025) 111711.DOI
- [6]L. Wang, C. Zeng, M. Zhang, J. Wu, L. Nie, Domain aware post training quantization for vision transformers in deployment, Pattern Recognition 171 (2026) 112182.DOI
- [7]T. Huang, S. You, F. Wang, C. Qian, C. Xu, DIST+: Knowledge distillation from a stronger adaptive teacher, IEEE Transactions on Pattern Analysis and Machine Intelligence 47 (2025) 5571–5585.DOI
- [8]M. Gabor, R. Zdunek, Reduced storage direct tensor ring decomposition for convolutional neural networks compression, Neural Networks 193 (2026) 107994.DOI
- [9]H. Cheng, M. Zhang, J. Q. Shi, A survey on deep neural network pruning: Taxonomy, comparison, analysis, and recommendations, IEEE Transactions on Pattern Analysis and Machine Intelligence 46 (2024) 10558–10578.DOI
- [10]J. Yan, W. Huang, L. Yin, W. Xie, Y. Zhao, Co-synthesizing multiple similar filters to enhance the feature diversity of pruned network, Pattern Recognition 179 (2026) 113948.DOI
- [11]Y. He, L. Xiao, Structured pruning for deep convolutional neural networks: A survey, IEEE Transactions on Pattern Analysis and Machine Intelligence 46 (2024) 2900–2919.DOI
- [12]Y. Lian, P. Peng, K. Jiang, W. Xu, Cross-layer importance evaluation for neural network pruning, Neural Networks 179 (2024) 106496.DOI
- [13]Y. Wang, S. Guo, J. Guo, J. Zhang, W. Zhang, C. Yan, Y. Zhang, Towards performance-maximizing neural network pruning via global channel attention, Neural Networks 171 (2024) 104–113.DOI
- [14]Z. Zhang, C. Cao, F. Zheng, T. Sun, L. Zhao, GCPNet: Gradient-aware channel pruning network with bilateral coupled sampling strategy, Expert Systems with Applications 266 (2025) 126104.DOI
- [15]M. Ma, T. Zhang, Z. Wang, Y. Wang, T. Du, W. Li, Global channel pruning with self-supervised mask learning, IEEE Transactions on Circuits and Systems for Video Technology 35 (2025) 2013–2025.DOI
- [16]M. Shi, C. Liu, J. Jiao, Q. Ye, Self-supervised feature-gate coupling for dynamic network pruning, Pattern Recognition 154 (2024) 110594.DOI
- [17]L. Yang, S. Gu, C. Shen, X. Zhao, Q. Hu, Soft independence guided filter pruning, Pattern Recognition 153 (2024) 110488.
- [18]V. T. Pham, Y. Zniyed, T. P. Nguyen, Efficient tensor decomposition-based filter pruning, Neural Networks 178 (2024) 106393.DOI
- [19]A. M. Shaikh, Y. Kang, A. Kumar, Y.-b. Zhao, Dynamic filter pruning via unified importance and redundancy, Information Sciences 729 (2026) 122845.DOI
- [20]E. Chen, H. Wang, Z. Shi, W. Zhang, Channel pruning for convolutional neural networks using l₀-norm constraints, Neurocomputing 636 (2025) 129925.DOI
- [21]L. Gonzalez-Carabarin, I. A. M. Huijben, B. Veeling, A. Schmid, R. J. G. van Sloun, Dynamic probabilistic pruning: A general framework for hardware-constrained pruning at different granularities, IEEE Transactions on Neural Networks and Learning Systems 35 (2024) 733–744.
- [22]H. Ye, B. Zhang, T. Chen, J. Fan, B. Wang, Performance-aware approximation of global channel pruning for multitask CNNs, IEEE Transactions on Pattern Analysis and Machine Intelligence 45 (2023) 10267–10284.arxiv.org/abs/2303.11923
- [23]V. T. Pham, Y. Zniyed, T. P. Nguyen, Enhanced network compression through tensor decompositions and pruning, IEEE Transactions on Neural Networks and Learning Systems 36 (2025) 4358–4370.DOI
- [24]M. Lin, R. Ji, Y. Wang, Y. Zhang, B. Zhang, Y. Tian, L. Shao, HRank: Filter pruning using high-rank feature map, in: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, IEEE, 2020, pp. 1526–1535.arxiv.org/abs/2002.10179
- [25]Y. Sui, M. Yin, Y. Xie, H. Phan, S. Aliari Zonouz, B. Yuan, CHIP: CHAnnel independence-based pruning for compact neural networks, in: M. Ranzato, A. Beygelzimer, Y. Dauphin, P. Liang, J. W. Vaughan (Eds.), Advances in Neural Information Processing Systems, volume 34, Curran Associates, Inc., 2021, pp. 24604–24616.arxiv.org/abs/2110.13981
- [26]S. Guo, L. Zhang, X. Zheng, Y. Wang, Y. Li, F. Chao, C. Wu, S. Zhang, R. Ji, Automatic network pruning via Hilbert-Schmidt independence criterion lasso under information bottleneck principle, in: 2023 IEEE/CVF International Conference on Computer Vision, IEEE, 2023, pp. 17412–17423.DOI
- [27]V. T. Pham, Y. Zniyed, T. P. Nguyen, Singular values-driven automated filter pruning, Neural Networks 192 (2025) 107857.DOI
- [28]G. Tian, Y. Sun, Y. Liu, X. Zeng, M. Wang, Y. Liu, J. Zhang, J. Chen, Adding before pruning: Sparse filter fusion for deep convolutional neural networks via auxiliary attention, IEEE Transactions on Neural Networks and Learning Systems 36 (2025) 3930–3942.
- [29]Y. Li, L. Wang, X. Tao, IPPKD: An iterative pruning method with prompt knowledge distillation, Pattern Recognition 180 (2026) 114640.DOI
- [30]S. Lin, R. Ji, C. Yan, B. Zhang, L. Cao, Q. Ye, F. Huang, D. Doermann, Towards optimal structured CNN pruning via generative adversarial learning, in: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition, IEEE, 2019, pp. 2785–2794.arxiv.org/abs/1903.09291
- [31]J. Zhang, S. Zhong, A. Ye, Z. Liu, S. Zhao, K. Zhou, L. Li, S.-H. Choi, R. Chen, X. Hu, S. Xu, V. Chaudhary, Flexible group count enables hassle-free structured pruning, in: 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition, IEEE, 2025, pp. 4807–4818.
- [32]Y. Wu, Y. Lan, L. Zhang, Y. Xiang, Feature flow regularization: Improving structured sparsity in deep neural networks, Neural Networks 161 (2023) 598–613.DOI
- [33]V. Boža, V. Macko, Two sparse matrices are better than one: Sparsifying neural networks with double sparse factorization, in: International Conference on Learning Representations, 2025.arxiv.org/abs/2409.18850
- [34]Z. Li, F. Nie, J. Bian, D. Wu, X. Li, Sparse PCA via ℓ₂,ₚ-norm regularization for unsupervised feature selection, IEEE Transactions on Pattern Analysis and Machine Intelligence 45 (2023) 5322–5328.
- [35]J. Liu, X. Ju, X. Xiu, W. Liu, Bi-level unsupervised feature selection with ℓ₂,₀-norm, Pattern Recognition 181 (2027) 114687.DOI
- [36]X. Xiu, C. Huang, P. Shang, W. Liu, Bi-sparse unsupervised feature selection, IEEE Transactions on Image Processing 34 (2025) 7407–7421.DOI
- [37]O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, L. Fei-Fei, ImageNet large scale visual recognition challenge, International Journal of Computer Vision 115 (2015) 211–252.
- [38]K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in: 2016 IEEE Conference on Computer Vision and Pattern Recognition, IEEE, 2016, pp. 770–778.arxiv.org/abs/1512.03385
- [39]J.-H. Luo, H. Zhang, H.-Y. Zhou, C.-W. Xie, J. Wu, W. Lin, ThiNet: Pruning CNN filters for a thinner net, IEEE Transactions on Pattern Analysis and Machine Intelligence 41 (2019) 2525–2538.DOI



