Pith. sign in

REVIEW 3 major objections 5 minor 4 cited by

FCOS: A Two-Stage Recoverable Model Pruning Framework for Automatic Modulation Recognition

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read FCOS is a two-stage pruning framework that claims to compress automatic modulation recognition networks by more than 95% in FLOPs and parameters while keeping accuracy nearly unchanged, and it demonstrates this on three datasets and three…

desk verdict A plausible two-stage pruning pipeline with a real experimental flaw: Table IV compares methods from different starting checkpoints, so the headline 'outperforms' claim is not yet supported. read the letter →

arxiv 2505.21571 v1 pith:AGNZ736A submitted 2025-05-27 cs.LG cs.AI

Yao Lu , Tengfei Ma , Zeyu Wang , Zhuangzhi Chen , Dongwei Xu , Yun Lin , Qi Xuan , Guan Gui This is my paper · ORCID
classification cs.LGcs.AI
keywords AutomaticModulationRecognitionModelPruningChannelLayerFusionLinearProbingHierarchicalClusteringCompression
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

FCOS is a pruning framework that asks whether a convolutional network for automatic modulation recognition can be compressed far beyond what channel or layer pruning alone allows, without a catastrophic accuracy drop. The paper's central claim is that it can: first cluster and average the most similar channels in each layer, then use linear probing to find layers that have collapsed under that compression and delete them. On Sig2019-12, FCOS removes 95.51% of FLOPs and 95.31% of parameters from ResNet56 while losing only 0.46% accuracy, and it beats five channel-pruning and four layer-pruning baselines across three datasets and three model architectures. A sympathetic reading is that the combination of fine-grained channel fusion and coarse-grained layer removal is what unlocks extreme compression.

What carries the argument

The load-bearing mechanism is the pairing of two complementary operations. First, average-linkage hierarchical clustering on cosine distances between output-channel weight vectors groups redundant channels, and the channels in each cluster are fused by simple averaging (Equation 5), which simultaneously slims the layer and averages out channel noise. Second, the Layer Collapse Diagnosis module computes a linear-probe accuracy $Acc_i$ for every layer's frozen features and applies the redundancy criterion $|Acc_i - Acc_{i-1}| \leq \beta$ (Equation 8); layers that pass the criterion are judged collapsed and removed. The linear probe is the instrument that detects when aggressive channel pruning has destroyed a layer's contribution, converting a width-pruning failure into a depth-pruning opportunity.

What would settle it

Run the LaCD stage on a pruned model, then for every layer it flags as collapsed measure the actual end-to-end accuracy change when that layer alone is removed and the model is fine-tuned; if many flagged layers produce large accuracy drops, the linear-probe criterion is not a faithful measure of layer importance.

Watch

Extended reading notes

Core claim

FCOS establishes that extreme compression of AMR models is achievable by a fine-to-coarse two-stage procedure. Stage one performs channel-level pruning by hierarchical clustering of output-channel weight vectors by cosine distance, then replaces each cluster with the average weight tensor, shrinking the network's width. Stage two, LaCD, freezes the pruned network, trains a linear classifier on each layer's features, and removes any layer whose probe accuracy differs from the previous layer's by at most a threshold $\beta$. The reported result is that this pipeline reduces ResNet56 on Sig2019-12 to 4.49% of its FLOPs and 4.69% of its parameters while keeping accuracy at 67.45% versus the original 67.91%, and similar or better trade-offs appear on RML2016.10a, RML2018.01a, CNN1D, and SigNet50.

Load-bearing premise

The weakest assumption is that a layer can be safely deleted whenever a linear classifier read from its outputs performs no better than the linear classifier read from the previous layer's outputs.

Editorial extensions

If this is right

  • On Sig2019-12, FCOS cuts ResNet56's FLOPs by 95.51% and parameters by 95.31% with only a 0.46% accuracy drop.
  • At high pruning rates where pure channel pruning collapses (for example, SFP drops 49.00% on Sig2019-12), FCOS keeps accuracy nearly unchanged.
  • Against layer-pruning baselines that achieve roughly 90% compression, FCOS reaches 95% or higher compression with a smaller accuracy loss.
  • The method generalizes across CNN1D, ResNet56, and SigNet50 on RML2016.10a, Sig2019-12, and RML2018.01a, and in several cases slightly improves accuracy after pruning.
  • Because collapsed-layer removal happens after massive channel pruning, the two-stage design is what makes extreme compression stable.

Reading between the lines

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

  • The linear-probe criterion is a generic collapse detector, so the same fine-to-coarse recipe could be applied to other signal-classification networks or even image models, provided the probe threshold is re-tuned.
  • The paper never reports the value of $\beta$ or a sensitivity analysis, so an immediate extension would test whether the method's advantage survives across a range of $\beta$ values; if it does not, the published results hinge on an unreported hyperparameter.
  • A natural stress test is to compare LaCD's diagnosis against leave-one-out layer ablation: if layers flagged as collapsed by linear probing are not the same layers whose removal costs the least end-to-end accuracy, the proxy is measuring something other than true layer importance.
  • The averaging-based fusion suggests that pruning and robustness are linked: fused channels may act like weight averaging, and measuring the pruned model's robustness to noise or distribution shift would clarify whether the accuracy gains come from regularization rather than compression alone.
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 FCOS, a two-stage pruning framework for automatic modulation recognition (AMR) models. The first stage performs channel pruning by hierarchically clustering channels based on weight similarity and fusing channels within each cluster via averaging. The second stage, Layer Collapse Diagnosis (LaCD), fine-tunes the pruned model briefly, then uses linear probing on frozen intermediate features to identify layers whose marginal classification accuracy improvement is below a threshold beta, and removes those layers. The authors claim that this fine-to-coarse strategy achieves extreme compression (e.g., 95.51% FLOPs and 95.31% parameter reduction on ResNet56 with Sig2019-12) while maintaining accuracy close to the original model, and that FCOS outperforms five channel-pruning and four layer-pruning baselines across three AMR datasets and three model architectures. The paper includes ablations on the similarity metric, fusion weighting, and fusion order, and provides a link to code.

Significance. If the central claims hold, FCOS would address a real limitation of channel pruning at high compression rates, where models are prone to structural collapse. The idea of combining channel fusion with linear-probing-based layer removal is interesting and the paper demonstrates a useful intuition: after aggressive channel pruning, some layers become redundant and can be removed without additional accuracy loss. The authors also provide code and carry out several focused ablations, which are strengths. However, the primary comparative claim—that FCOS outperforms existing channel and layer pruning methods—is currently not supported by the evidence as presented, because the baseline comparisons are not controlled. The core issue is that Table IV reports different original accuracies for the same model and dataset across methods, so the reported accuracy differences reflect the starting checkpoints as much as the pruning algorithms. This is a load-bearing problem for a paper whose headline result is a performance comparison.

major comments (3)
  1. [Table IV] The central comparison against channel-pruning baselines is uncontrolled because the 'Original Acc' column differs across methods for the same model and dataset. For example, for ResNet56 on RML2018.01a, the channel-pruning baselines start from 92.00%, 91.95%, 92.56%, 91.95%, and 92.00%, whereas FCOS and the layer-pruning baselines start from 89.70%. Similarly, on RML2016.10a the channel baselines start from 60.70%–62.63% while FCOS starts from 62.06%, and on Sig2019-12 the channel baselines start from 66.20%–67.16% while FCOS starts from 67.91%. Since the reported ΔAcc values are computed relative to these different starting accuracies, the claim in the abstract and Section IV-D that FCOS 'outperforms' the five channel-pruning methods is not supported by this table. The authors should rerun all baselines from the same pre-trained checkpoints used for FCOS, or otherwise ensure that the original accuracy is identical for all methods on a given model and dataset, and then report the resulting ΔAcc values.
  2. [Eq. (8), Section III-B] The layer redundancy threshold beta, which is the key parameter of the LaCD module, is never reported anywhere in the paper. Equation (8) defines layer removal by |Acc_i - Acc_{i-1}| <= beta, but no values of beta are given for any of the experiments, and no sensitivity analysis is provided. Without this information, the second stage of FCOS is not reproducible, and the reader cannot tell how many layers are removed or how sensitive the final accuracy is to the choice of beta. This is particularly important because beta directly controls the trade-off between compression and accuracy in the layer-pruning stage.
  3. [Section III-B, Eq. (7)] The LaCD method assumes that the linear probe accuracy difference |Acc_i - Acc_{i-1}| is a faithful proxy for the contribution of layer i to the end-to-end model, but this assumption is not validated. The paper provides no comparison of LaCD's layer selection against a ground-truth measure of layer importance (e.g., leave-one-out accuracy) and no evidence that the layers flagged as collapsed are indeed the least important. If the linear probe over- or under-estimates layer importance, the method could remove important layers or fail to remove truly collapsed ones; the current experiments cannot distinguish these cases. The authors should provide a direct analysis of the linear probe's predictive validity for layer removal, or at minimum report the probe accuracies Acc_i for the pruned models and show that the removed layers have small marginal contributions.
minor comments (5)
  1. [Section IV] The text says 'we verify the effectiveness of FCOS on 4 models across 3 datasets,' but the experiments use only three models (ResNet56, CNN1D, and SigNet50). This should be corrected to '3 models'.
  2. [Table IV] The column headers use 'PR' ambiguously for both 'pruning rate' and 'params PR'/'FLOPs PR'; consider renaming to 'PR (%)' or using separate clear labels to avoid confusion.
  3. [Section III-A, Eq. (3)] The per-layer channel pruning rate epsilon is a key hyperparameter, but the paper does not report the specific epsilon values used to achieve the reported pruning rates in Table IV. Reporting these values would improve reproducibility.
  4. [Section IV-F] The text says the ablation demonstrates that 'cosine similarity can be replaced by any similarity metric,' but only Euclidean similarity is tested. The statement should be softened to 'can be replaced by other similarity metrics, such as Euclidean similarity' to match the evidence.
  5. [Table II] In Table II, the 'Original Params' and 'Original FLOPs' columns are identical for the 'Normal' and 'Abnormal' rows, and the table is not referenced in the main text with an explanation of how the two rows were obtained; a sentence clarifying the experimental setup would help.

Circularity Check

1 steps flagged · score 2.0 of 10

Mild definitional circularity in the LaCD layer-collapse criterion; the core pruning results remain empirical and are not circular.

  1. self definitional [Section III-B, Eq. (8), and Algorithm 2, lines 8-10]
    "By analyzing the layer-level features in the frozen state, the layers that have little impact on the overall performance can be identified at a lower computational cost and incorporated into Pr. Here, we define the impact of the i-th layer on the model performance as follows: |Acc_i - Acc_{i-1}| <= beta, where beta is a predefined layer redundancy threshold. Finally, we remove collapsed layers in Pr."

    Eq. (8) defines a layer's 'impact' as the linear-probe accuracy difference, and Algorithm 2 marks layer i for removal exactly when |Acc_i - Acc_{i-1}| <= beta. Hence LaCD's 'identification of collapsed layers' is a restatement of the threshold condition; the removal decision is forced by the definition. The paper supplies no independent derivation of this proxy and does not report beta, so the layer-selection step has no content beyond the operational definition. Final accuracies are measured externally, so the main empirical claims are not circular.

full rationale

The central claims of FCOS are empirical: 95.51% FLOPs reduction, 95.31% parameter reduction, and measured accuracy after pruning on three AMR datasets. No equation fits a parameter and then renames it as a prediction; the channel-fusion equations (Eqs. 1-5) define a clustering and averaging procedure whose outcome is evaluated on held-out accuracy. The only near-circular element is the Layer Collapse Diagnosis step, where 'collapse' is defined by the same linear-probe accuracy-difference threshold used to remove layers; this is a definitional heuristic rather than a derived result. Self-citations, such as Refs. [33], [34], [37], and [38], are prior layer-pruning works used as related work or baselines and are not load-bearing for the method's derivation. The uncontrolled comparison in Table IV, where Original Acc values differ for the same model and dataset, is a genuine correctness risk, but it is an experimental-control issue rather than circularity. Overall, the paper's main contribution is an empirically evaluated pipeline with only mild definitional circularity.

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

The central claim rests on a set of empirical hyperparameters (epsilon, beta) and on assumptions about weight similarity, linear-probe fidelity, and fine-tuning recovery. None of these are derived from first principles, and the paper does not report the values of the two most important hyperparameters. There are no newly invented physical entities; the method operates entirely within standard neural network components.

free parameters (5)
  • epsilon = not reported
    Channel pruning rate used in Eq. (3) to set the number of clusters per layer. Its value is not given in the paper, yet it directly determines the degree of channel compression and hence the reported FLOPs and parameter reductions.
  • beta = not reported
    Layer redundancy threshold in Eq. (8). The paper describes it as 'predefined' but never states its value or how it was chosen, so the number of removed layers is not reproducible.
  • fine-tuning epochs = 20 (channel pruning), 5 (linear probe), 80 (final)
    The training schedule is specified in Section IV-B, but these are hand-selected hyperparameters that affect all results.
  • learning rate and batch size = 0.001 and 128
    Used in all fine-tuning and probing stages; standard but arbitrary choices that could influence the reported accuracy.
  • hierarchical clustering linkage = average linkage
    The paper chooses average linkage from the three standard options without a sensitivity study, so the clustering outcome depends on this choice.
assumptions (5)
  • domain assumption Learned weights of a pre-trained model encode capabilities that can be exploited by averaging similar channels.
    Invoked in Section III-A citing [52] to justify that weight similarity is a proxy for redundancy and that averaging is safe.
  • domain assumption Cosine similarity between channel weight vectors identifies channels that can be fused without harming the network.
    Used in Eq. (1)-(2) as the basis for clustering. The paper tests one alternative metric in Section IV-F but does not prove the assumption.
  • domain assumption Linear probe accuracy on a layer's output is a faithful proxy for the layer's contribution to end-to-end accuracy.
    This is the core of LaCD in Section III-B; the paper relies on it to decide when a layer is 'collapsed' and removable.
  • domain assumption Fine-tuning after pruning can restore performance close to the original model.
    The entire pipeline depends on fine-tuning (20 epochs, then 80 epochs) to recover accuracy; no theory guarantees this for extreme pruning rates.
  • standard math Hierarchical clustering with average linkage produces groups of channels whose average is a valid representative.
    Relies on the standard hierarchical clustering algorithm [60]; the averaging in Eq. (5) is a simple arithmetic mean.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FCOS: A Two-Stage Recoverable Model Pruning Framework for Automatic Modulation Recognition." pith.science (2026). https://pith.science/paper/AGNZ736A

@misc{pith2026250521571,
  author       = {Pith},
  title        = {Pith review of: FCOS: A Two-Stage Recoverable Model Pruning Framework for Automatic Modulation Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AGNZ736A}},
  note         = {Machine review of arXiv:2505.21571}
}
read the original abstract

With the rapid development of wireless communications and the growing complexity of digital modulation schemes, traditional manual modulation recognition methods struggle to extract reliable signal features and meet real-time requirements in modern scenarios. Recently, deep learning based Automatic Modulation Recognition (AMR) approaches have greatly improved classification accuracy. However, their large model sizes and high computational demands hinder deployment on resource-constrained devices. Model pruning provides a general approach to reduce model complexity, but existing weight, channel, and layer pruning techniques each present a trade-off between compression rate, hardware acceleration, and accuracy preservation. To this end, in this paper, we introduce FCOS, a novel Fine-to-COarse two-Stage pruning framework that combines channel-level pruning with layer-level collapse diagnosis to achieve extreme compression, high performance and efficient inference. In the first stage of FCOS, hierarchical clustering and parameter fusion are applied to channel weights to achieve channel-level pruning. Then a Layer Collapse Diagnosis (LaCD) module uses linear probing to identify layer collapse and removes the collapsed layers due to high channel compression ratio. Experiments on multiple AMR benchmarks demonstrate that FCOS outperforms existing channel and layer pruning methods. Specifically, FCOS achieves 95.51% FLOPs reduction and 95.31% parameter reduction while still maintaining performance close to the original ResNet56, with only a 0.46% drop in accuracy on Sig2019-12. Code is available at https://github.com/yaolu-zjut/FCOS.

Figures

Figures reproduced from arXiv: 2505.21571 by the authors.

Figure 1
Figure 1. The pipeline of FCOS. In the first stage, hierarchical clustering and parameter fusion are applied to channel weights to achieve channel-level pruning. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Visualization of the changes in layers and channels during pruning of CNN1D on RML2016.10a. [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Comparison of the accuracy curve of the model obtained by FCOS [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 4 Pith papers

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

  1. A Multimodal Deep Learning Framework for Early Diagnosis of Liver Cancer via Optimized BiLSTM-AM-VMD Architecture

    cs.LG 2025-09 reject novelty 3.0 of 10

    The paper claims a BiLSTM-AM-VMD model achieves AUC 0.963 for early HCC diagnosis, but the evidence is undermined by contradictory dataset descriptions and missing artifacts.

  2. Multi-Modal Machine Learning Framework for Predicting Early Recurrence of Brain Tumors Using MRI and Clinical Biomarkers

    cs.LG 2025-09 reject novelty 3.0 of 10

    XGBoost combining MRI radiomics and clinical biomarkers reportedly reaches C-index 0.782 for early brain tumor recurrence, but the paper's methods describe a liver-cancer cohort and no evaluation of its claimed tempor...

  3. DUSE: A Data Expansion Framework for Low-resource Automatic Modulation Recognition based on Active Learning

    cs.LG 2025-07 conditional novelty 3.0 of 10

    DUSE selects top-margin samples from an auxiliary radio-signal pool via iterative active learning to expand low-resource AMR training sets.

  4. ReStNet: A Reusable & Stitchable Network for Dynamic Adaptation on IoT Devices

    cs.CV 2025-06 reject novelty 3.0 of 10

    ReStNet builds a hybrid model from two pre-trained networks by stitching at the most CKA-similar layer and fine-tuning only the stitching layer, claiming flexible runtime trade-offs for IoT devices.

Reference graph

Works this paper leans on

68 extracted references · 36 canonical work pages · cited by 4 Pith papers

  1. [1]

    Type of modulation identification using wavelet transform and neural network,

    M. Walenczykowska and A. Kawalec, “Type of modulation identification using wavelet transform and neural network,” Bulletin of the Polish Academy of Sciences. Technical Sciences , vol. 64, no. 1, pp. 257–261, 2016

    2016

  2. [2]

    Wavelet transform based modula- tion classification for 5g and uav communication in multipath fading channel,

    W. Li, Z. Dou, L. Qi, and C. Shi, “Wavelet transform based modula- tion classification for 5g and uav communication in multipath fading channel,” Physical Communication, vol. 34, pp. 272–282, 2019

    2019

  3. [3]

    Phasma: An automatic modulation classification system based on random forest,

    K. Triantafyllakis, M. Surligas, G. Vardakis, and S. Papadakis, “Phasma: An automatic modulation classification system based on random forest,” in 2017 IEEE International Symposium on Dynamic Spectrum Access Networks (DySPAN). IEEE, 2017, pp. 1–3

    2017

  4. [4]

    Cyclic spectral analysis of ofdm/oqam signals,

    D. Vu ˇci´c, S. Vukoti ´c, and M. Eri ´c, “Cyclic spectral analysis of ofdm/oqam signals,” AEU-International Journal of Electronics and Communications, vol. 73, pp. 139–143, 2017

    2017

  5. [5]

    Automatic mod- ulation classification based on high order cumulants and hierarchical polynomial classifiers,

    A. Abdelmutalab, K. Assaleh, and M. El-Tarhuni, “Automatic mod- ulation classification based on high order cumulants and hierarchical polynomial classifiers,” Physical Communication , vol. 21, pp. 10–18, 2016

    2016

  6. [6]

    Over-the-air deep learning based radio signal classification,

    T. J. O’Shea, T. Roy, and T. C. Clancy, “Over-the-air deep learning based radio signal classification,” IEEE Journal of Selected Topics in Signal Processing, vol. 12, no. 1, pp. 168–179, 2018

    2018

  7. [7]

    Convolutional radio mod- ulation recognition networks,

    T. J. O’Shea, J. Corgan, and T. C. Clancy, “Convolutional radio mod- ulation recognition networks,” in Engineering Applications of Neural Networks: 17th International Conference, EANN 2016, Aberdeen, UK, September 2-5, 2016, Proceedings 17 . Springer, 2016, pp. 213–226

    2016

  8. [8]

    An improved neural network pruning technology for automatic modulation classification in edge devices,

    Y . Lin, Y . Tu, and Z. Dou, “An improved neural network pruning technology for automatic modulation classification in edge devices,” IEEE Transactions on Vehicular Technology , vol. 69, no. 5, pp. 5703– 5706, 2020

    2020

Show all 68 references
  1. [9]

    Signet: A novel deep learning framework for radio signal classification,

    Z. Chen, H. Cui, J. Xiang, K. Qiu, L. Huang, S. Zheng, S. Chen, Q. Xuan, and X. Yang, “Signet: A novel deep learning framework for radio signal classification,” IEEE Transactions on Cognitive Communi- cations and Networking , vol. 8, no. 2, pp. 529–541, 2021. JOURNAL OF LATEX ...

    2021

  2. [10]

    Contour stella image and deep learning for signal recognition in the physical layer,

    Y . Lin, Y . Tu, Z. Dou, L. Chen, and S. Mao, “Contour stella image and deep learning for signal recognition in the physical layer,” IEEE Transactions on Cognitive Communications and Networking , vol. 7, no. 1, pp. 34–46, 2020

    2020

  3. [11]

    Complex-valued networks for automatic modulation classification,

    Y . Tu, Y . Lin, C. Hou, and S. Mao, “Complex-valued networks for automatic modulation classification,” IEEE Transactions on Vehicular Technology, vol. 69, no. 9, pp. 10 085–10 089, 2020

    2020

  4. [12]

    Adversarial attacks in modulation recognition with convolutional neural networks,

    Y . Lin, H. Zhao, X. Ma, Y . Tu, and M. Wang, “Adversarial attacks in modulation recognition with convolutional neural networks,” IEEE Transactions on Reliability , vol. 70, no. 1, pp. 389–401, 2020

    2020

  5. [13]

    Lightweight automatic modulation classification via progres- sive differentiable architecture search,

    X. Zhang, X. Chen, Y . Wang, G. Gui, B. Adebisi, H. Sari, and F. Adachi, “Lightweight automatic modulation classification via progres- sive differentiable architecture search,” IEEE Transactions on Cognitive Communications and Networking , 2023

    2023

  6. [14]

    Multi-view discriminant framework for automatic modulation open set recognition,

    J. Hou, D. Xu, F. Song, Z. Chen, Q. Xuan, S. Zheng, Y . Lin, and X. Yang, “Multi-view discriminant framework for automatic modulation open set recognition,” IEEE Transactions on Communications , 2024

    2024

  7. [15]

    Mclrl: A multi- domain contrastive learning with reinforcement learning framework for few-shot modulation recognition,

    D. Xu, Y . Zhu, Y . Lu, Y . Feng, Y . Lin, and Q. Xuan, “Mclrl: A multi- domain contrastive learning with reinforcement learning framework for few-shot modulation recognition,” arXiv preprint arXiv:2502.19071, 2025

    2025 arXiv

  8. [16]

    Data-driven deep learning for automatic modulation recognition in cognitive radios,

    Y . Wang, M. Liu, J. Yang, and G. Gui, “Data-driven deep learning for automatic modulation recognition in cognitive radios,” IEEE Transac- tions on Vehicular Technology, vol. 68, no. 4, pp. 4074–4077, 2019

    2019

  9. [17]

    Automatic modulation recognition for spectrum sensing using nonuniform compressive samples,

    C. W. Lim and M. B. Wakin, “Automatic modulation recognition for spectrum sensing using nonuniform compressive samples,” in2012 IEEE International Conference on Communications (ICC) . IEEE, 2012, pp. 3505–3510

    2012

  10. [18]

    A feature weighted hybrid ica-svm approach to automatic modulation recognition,

    D. Boutte and B. Santhanam, “A feature weighted hybrid ica-svm approach to automatic modulation recognition,” in 2009 IEEE 13th Digital Signal Processing Workshop and 5th IEEE Signal Processing Education Workshop. IEEE, 2009, pp. 399–403

    2009

  11. [19]

    Automatic modulation recognition of unknown interference signals based on graph model,

    Q. Zhang, H. Ji, L. Li, and Z. Zhu, “Automatic modulation recognition of unknown interference signals based on graph model,” IEEE Wireless Communications Letters, 2024

    2024

  12. [20]

    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

    2014 arXiv

  13. [21]

    Amc-net: An effective network for automatic modulation classification,

    J. Zhang, T. Wang, Z. Feng, and S. Yang, “Amc-net: An effective network for automatic modulation classification,” in ICASSP 2023- 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  14. [22]

    Efficient automatic modulation classification in non-terrestrial networks with snn-based transformer,

    D. Zeng, Y . Xiao, W. Liu, H. Du, E. Zhang, D. Zhang, Y . Wang, M. Zhang, and W. Chen, “Efficient automatic modulation classification in non-terrestrial networks with snn-based transformer,” IEEE Internet of Things Journal , 2024

    2024

  15. [23]

    Towards building a high-performance intelligent radio network through deep learning: Addressing data privacy, adver- sarial robustness, network structure, and latency requirements

    A. S. M. M. Jameel, “Towards building a high-performance intelligent radio network through deep learning: Addressing data privacy, adver- sarial robustness, network structure, and latency requirements.” Ph.D. dissertation, Purdue University Graduate School, 2024

    2024

  16. [24]

    Edge learning for b5g networks with distributed signal processing: Semantic communication, edge computing, and wireless sensing,

    W. Xu, Z. Yang, D. W. K. Ng, M. Levorato, Y . C. Eldar, and M. Debbah, “Edge learning for b5g networks with distributed signal processing: Semantic communication, edge computing, and wireless sensing,” IEEE journal of selected topics in signal processing , vol. 17, no. 1, pp. ...

    2023

  17. [25]

    Task- oriented communications for 6g: Vision, principles, and technologies,

    Y . Shi, Y . Zhou, D. Wen, Y . Wu, C. Jiang, and K. B. Letaief, “Task- oriented communications for 6g: Vision, principles, and technologies,” IEEE Wireless Communications, vol. 30, no. 3, pp. 78–85, 2023

    2023

  18. [26]

    A systematic dnn weight pruning framework using alternating direction method of multipliers,

    T. Zhang, S. Ye, K. Zhang, J. Tang, W. Wen, M. Fardad, and Y . Wang, “A systematic dnn weight pruning framework using alternating direction method of multipliers,” in ECCV, 2018, pp. 184–199

    2018

  19. [27]

    Pconv: The missing but desirable sparsity in dnn weight pruning for real-time execution on mobile devices,

    X. Ma, F.-M. Guo, W. Niu, X. Lin, J. Tang, K. Ma, B. Ren, and Y . Wang, “Pconv: The missing but desirable sparsity in dnn weight pruning for real-time execution on mobile devices,” in AAAI, vol. 34, no. 04, 2020, pp. 5117–5124

    2020

  20. [28]

    Combining weight pruning and knowledge distillation for cnn compression,

    N. Aghli and E. Ribeiro, “Combining weight pruning and knowledge distillation for cnn compression,” in CVPR, 2021, pp. 3191–3198

    2021

  21. [29]

    Pruning filter in filter,

    F. Meng, H. Cheng, K. Li, H. Luo, X. Guo, G. Lu, and X. Sun, “Pruning filter in filter,” NeurIPS, vol. 33, pp. 17 629–17 640, 2020

    2020

  22. [30]

    Importance estimation for neural network pruning,

    P. Molchanov, A. Mallya, S. Tyree, I. Frosio, and J. Kautz, “Importance estimation for neural network pruning,” in CVPR, 2019, pp. 11 264– 11 272

    2019

  23. [31]

    Channel pruning for accelerating very deep neural networks,

    Y . He, X. Zhang, and J. Sun, “Channel pruning for accelerating very deep neural networks,” in ICCV, 2017, pp. 1389–1397

    2017

  24. [32]

    Discrimination-aware channel pruning for deep neural networks,

    Z. Zhuang, M. Tan, B. Zhuang, J. Liu, Y . Guo, Q. Wu, J. Huang, and J. Zhu, “Discrimination-aware channel pruning for deep neural networks,” Advances in neural information processing systems , vol. 31, 2018

    2018

  25. [33]

    Understanding the dynamics of dnns using graph modularity,

    Y . Lu, W. Yang, Y . Zhang, Z. Chen, J. Chen, Q. Xuan, Z. Wang, and X. Yang, “Understanding the dynamics of dnns using graph modularity,” in European Conference on Computer Vision. Springer, 2022, pp. 225– 242

    2022

  26. [34]

    Sr-init: An interpretable layer pruning method,

    H. Tang, Y . Lu, and Q. Xuan, “Sr-init: An interpretable layer pruning method,” in ICASSP. IEEE, 2023, pp. 1–5

    2023

  27. [35]

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

    S. Chen and Q. Zhao, “Shallowing deep networks: Layer-wise pruning based on feature representations,” TPAMI, vol. 41, no. 12, pp. 3048– 3056, 2018

    2018

  28. [36]

    Data-driven sparse structure selection for deep neural networks,

    Z. Huang and N. Wang, “Data-driven sparse structure selection for deep neural networks,” in ECCV, 2018, pp. 304–320

    2018

  29. [37]

    A generic layer pruning method for signal modulation recognition deep learning models,

    Y . Lu, Y . Zhu, Y . Li, D. Xu, Y . Lin, Q. Xuan, and X. Yang, “A generic layer pruning method for signal modulation recognition deep learning models,” IEEE Transactions on Cognitive Communications and Networking, 2024

    2024

  30. [38]

    Reassessing layer pruning in llms: New insights and methods,

    Y . Lu, H. Cheng, Y . Fang, Z. Wang, J. Wei, D. Xu, Q. Xuan, X. Yang, and Z. Zhu, “Reassessing layer pruning in llms: New insights and methods,” arXiv preprint arXiv:2411.15558 , 2024

    2024 arXiv

  31. [39]

    Rgp: Neural network pruning through regular graph with edges swap- ping,

    Z. Chen, J. Xiang, Y . Lu, Q. Xuan, Z. Wang, G. Chen, and X. Yang, “Rgp: Neural network pruning through regular graph with edges swap- ping,” IEEE Transactions on Neural Networks and Learning Systems , 2023

    2023

  32. [40]

    Radio machine learning dataset generation with gnu radio,

    T. J. O’shea and N. West, “Radio machine learning dataset generation with gnu radio,” in Proceedings of the GNU Radio Conference , vol. 1, no. 1, 2016

    2016

  33. [41]

    Filter pruning via measuring feature map information,

    L. Shao, H. Zuo, J. Zhang, Z. Xu, J. Yao, Z. Wang, and H. Li, “Filter pruning via measuring feature map information,”Sensors, vol. 21, no. 19, p. 6601, 2021

    2021

  34. [42]

    Filter pruning via geometric median for deep convolutional neural networks acceleration,

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

    2019

  35. [43]

    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

  36. [44]

    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,” in Proceedings of the 27th International Joint Conference on Artificial Intelligence , 2018, pp. 2234–2240

    2018

  37. [45]

    Channel pruning method for signal modulation recognition deep learning models,

    Z. Chen, Z. Wang, X. Gao, J. Zhou, D. Xu, S. Zheng, Q. Xuan, and X. Yang, “Channel pruning method for signal modulation recognition deep learning models,” IEEE Transactions on Cognitive Communica- tions and Networking , 2023

    2023

  38. [46]

    Dbp: Discrimination based block-level pruning for deep model acceleration,

    W. Wang, S. Zhao, M. Chen, J. Hu, D. Cai, and H. Liu, “Dbp: Discrimination based block-level pruning for deep model acceleration,” arXiv preprint arXiv:1912.10178 , 2019

    1912 arXiv

  39. [47]

    Lightweight automatic modulation classification based on decentralized learning,

    X. Fu, G. Gui, Y . Wang, T. Ohtsuki, B. Adebisi, H. Gacanin, and F. Adachi, “Lightweight automatic modulation classification based on decentralized learning,” IEEE Transactions on Cognitive Communica- tions and Networking , vol. 8, no. 1, pp. 57–70, 2021

    2021

  40. [48]

    Nas-amr: Neural architecture search-based automatic mod- ulation recognition for integrated sensing and communication systems,

    X. Zhang, H. Zhao, H. Zhu, B. Adebisi, G. Gui, H. Gacanin, and F. Adachi, “Nas-amr: Neural architecture search-based automatic mod- ulation recognition for integrated sensing and communication systems,” IEEE Transactions on Cognitive Communications and Networking , vol. 8, no....

    2022

  41. [49]

    Ultra lite convolutional neural network for automatic modulation classification in internet of unmanned aerial vehicles,

    L. Guo, Y . Wang, Y . Liu, Y . Lin, H. Zhao, and G. Gui, “Ultra lite convolutional neural network for automatic modulation classification in internet of unmanned aerial vehicles,” IEEE Internet of Things Journal , 2024

    2024

  42. [50]

    Pruning neural networks without any data by iteratively conserving synaptic flow,

    H. Tanaka, D. Kunin, D. L. Yamins, and S. Ganguli, “Pruning neural networks without any data by iteratively conserving synaptic flow,” Advances in neural information processing systems , vol. 33, pp. 6377– 6389, 2020

    2020

  43. [51]

    Deep model fusion: A survey,

    W. Li, Y . Peng, M. Zhang, L. Ding, H. Hu, and L. Shen, “Deep model fusion: A survey,” arXiv preprint arXiv:2309.15698 , 2023

    2023 arXiv

  44. [52]

    Merging models with fisher-weighted averaging,

    M. S. Matena and C. A. Raffel, “Merging models with fisher-weighted averaging,” Advances in Neural Information Processing Systems, vol. 35, pp. 17 703–17 716, 2022

    2022

  45. [53]

    Dataless knowl- edge fusion by merging weights of language models,

    X. Jin, X. Ren, D. Preotiuc-Pietro, and P. Cheng, “Dataless knowl- edge fusion by merging weights of language models,” arXiv preprint arXiv:2212.09849, 2022

    2022 arXiv

  46. [54]

    Population parameter averaging (papa),

    A. Jolicoeur-Martineau, E. Gervais, K. Fatras, Y . Zhang, and S. Lacoste- Julien, “Population parameter averaging (papa),” arXiv preprint arXiv:2304.03094, 2023

    2023 arXiv

  47. [55]

    Model ratatouille: Recycling diverse models for out-of-distribution gen- eralization,

    A. Ram ´e, K. Ahuja, J. Zhang, M. Cord, L. Bottou, and D. Lopez-Paz, “Model ratatouille: Recycling diverse models for out-of-distribution gen- eralization,” in International Conference on Machine Learning. PMLR, 2023, pp. 28 656–28 679. JOURNAL OF LATEX CLASS FILES, VOL. 14, N...

    2023

  48. [56]

    Faster cnns with direct sparse convolutions and guided pruning,

    J. Park, S. Li, W. Wen, P. T. P. Tang, H. Li, Y . Chen, and P. Dubey, “Faster cnns with direct sparse convolutions and guided pruning,” in ICLR, 2016

    2016

  49. [57]

    Eie: Efficient inference engine on compressed deep neural network,

    S. Han, X. Liu, H. Mao, J. Pu, A. Pedram, M. A. Horowitz, and W. J. Dally, “Eie: Efficient inference engine on compressed deep neural network,” ACM SIGARCH Computer Architecture News , vol. 44, no. 3, pp. 243–254, 2016

    2016

  50. [58]

    Robust statistics on riemannian manifolds via the geometric median,

    P. T. Fletcher, S. Venkatasubramanian, and S. Joshi, “Robust statistics on riemannian manifolds via the geometric median,” in 2008 IEEE Conference on Computer Vision and Pattern Recognition . IEEE, 2008, pp. 1–8

    2008

  51. [59]

    To filter prune, or to layer prune, that is the question,

    S. Elkerdawy, M. Elhoushi, A. Singh, H. Zhang, and N. Ray, “To filter prune, or to layer prune, that is the question,” in Proceedings of the Asian Conference on Computer Vision , 2020

    2020

  52. [60]

    Hierarchical clustering schemes,

    S. C. Johnson, “Hierarchical clustering schemes,” Psychometrika, vol. 32, no. 3, pp. 241–254, 1967

    1967

  53. [61]

    Emr- merging: Tuning-free high-performance model merging,

    C. Huang, P. Ye, T. Chen, T. He, X. Yue, and W. Ouyang, “Emr- merging: Tuning-free high-performance model merging,” Advances in Neural Information Processing Systems , vol. 37, pp. 122 741–122 769, 2024

    2024

  54. [62]

    Ties- merging: Resolving interference when merging models,

    P. Yadav, D. Tam, L. Choshen, C. A. Raffel, and M. Bansal, “Ties- merging: Resolving interference when merging models,” Advances in Neural Information Processing Systems , vol. 36, pp. 7093–7115, 2023

    2023

  55. [63]

    Averaging weights leads to wider optima and better generalization,

    P. Izmailov, D. Podoprikhin, T. Garipov, D. Vetrov, and A. G. Wilson, “Averaging weights leads to wider optima and better generalization,” arXiv preprint arXiv:1803.05407 , 2018

    2018 arXiv

  56. [64]

    Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time,

    M. Wortsman, G. Ilharco, S. Y . Gadre, R. Roelofs, R. Gontijo-Lopes, A. S. Morcos, H. Namkoong, A. Farhadi, Y . Carmon, S. Kornblith et al. , “Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time,” in International co...

    2022

  57. [65]

    Revisiting checkpoint aver- aging for neural machine translation,

    Y . Gao, C. Herold, Z. Yang, and H. Ney, “Revisiting checkpoint aver- aging for neural machine translation,” arXiv preprint arXiv:2210.11803, 2022

    2022 arXiv

  58. [66]

    Seasoning model soups for robustness to adversarial and natural distribution shifts,

    F. Croce, S.-A. Rebuffi, E. Shelhamer, and S. Gowal, “Seasoning model soups for robustness to adversarial and natural distribution shifts,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 12 313–12 323

    2023

  59. [67]

    Gnu radio: tools for exploring the radio frequency spec- trum,

    E. Blossom, “Gnu radio: tools for exploring the radio frequency spec- trum,” Linux journal, vol. 2004, no. 122, p. 4, 2004

    2004

  60. [68]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 770–778. Yao Lu received his B.S. degree from Zhejiang University of Technology and is currently pu...

    2016

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.