Pith. sign in

REVIEW 7 major objections 6 minor 2 cited by

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

T0 review · 7 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read DUSE, a data-expansion framework for low-resource automatic modulation recognition, selects the most uncertain samples from a large auxiliary signal dataset and iteratively retrains the selector, reporting accuracy gains over eight…

desk verdict A well-run study of margin selection plus active learning whose abstract overclaims consistency contradicted by its own tables. read the letter →

arxiv 2507.12011 v1 pith:23UXQTRK submitted 2025-07-16 cs.LG cs.AI

Yao Lu , Hongyu Gao , Zhuangzhi Chen , Dongwei Xu , Yun Lin , Qi Xuan , Guan Gui This is my paper · ORCID
classification cs.LGcs.AI
keywords AutomaticModulationRecognitionDataExpansionCoresetSelectionActiveLearningUncertaintyScoringLow-resourceSignalclassification
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

This paper tries to establish that a data-expansion framework called DUSE can relieve low-resource automatic modulation recognition by pulling the most useful samples out of a large related signal dataset, instead of collecting or labeling more data or synthesizing variants. The mechanism is an uncertainty score equal to the gap between a model's top two predicted class probabilities, paired with an active-learning loop that retrains the scoring model after each round of selected samples. The paper reports experiments on three AMR benchmarks under class-balanced and class-imbalanced conditions, comparing against eight coreset-selection baselines and claiming better accuracy in most settings plus strong cross-architecture transfer. If this is right, a practitioner with a small target dataset could use a large labeled AMR benchmark as raw material for training without new data collection.

What carries the argument

The central object is the dynamic uncertainty-scoring loop. The scorer is the margin $u = p_1^* - p_2^*$ between the two largest softmax probabilities; small values mark samples the current model finds most confusable, i.e., near decision boundaries. The loop retrains the model on the growing target set before each new round of scoring, so the definition of 'informative' adapts as the data distribution changes, and each round removes selected samples from the auxiliary pool to avoid re-selection.

What would settle it

Two checks would settle whether 'consistently outperforms' holds. The first is already in the paper: at the 1% expansion budget on Sig2019-12, Herding scores $38.33 \pm 12.0$ against DUSE's $20.98 \pm 0.27$, and on class-balanced RML2016.10a at 1%, DeepFool scores $54.61 \pm 0.75$ against DUSE's $52.15 \pm 0.48$; these are explicit counterexamples to a strict reading of the abstract's claim. The second is a domain-shift test the paper does not run: use an auxiliary pool recorded under a different channel, receiver, or SNR range while keeping the label set, and check whether DUSE still beats the baselines; if it does not, the motivating 'related but different field' scenario is unsupported.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that iterative margin-based uncertainty selection is an effective data-expansion strategy for AMR. Given a small labeled target set $D_T$ and a large labeled auxiliary set $D_A$, DUSE scores every auxiliary sample by $u = p_1^* - p_2^*$, the distance between the top two softmax probabilities under a model trained on the current target set, then moves the top-$K$ most uncertain samples into $D_T$, removes them from $D_A$, retrains the model on the enlarged set, and repeats for $R$ rounds. The paper argues that this loop selects samples near decision boundaries, keeps the augmented set closer to the target distribution than the baselines do, and that this is why accuracy rises with expansion budget and why the selected set still works well when evaluated on architectures other than the one used for scoring.

Load-bearing premise

The load-bearing premise is that the auxiliary dataset and the target dataset are distributionally equivalent: in every experiment the target is a 1% class-balanced random slice of the full training set and the auxiliary pool is the remaining 99% with identical class proportions, so a genuinely different auxiliary domain would be untested.

Editorial extensions

If this is right

  • DUSE can be used as a drop-in expansion step: with only a small labeled target set and any large labeled AMR dataset, users can add $K$ selected samples per round without changing the downstream training procedure.
  • The gains grow with budget in the class-imbalanced setting on RML2016.10a, moving from 48.68 at 1% to 80.69 at 19%, so the method is most useful when more auxiliary data can be afforded.
  • Because the expanded set transfers across architectures, one selection performed with a 2D-CNN can serve other models; on Sig2019-12, SigNet reaches 93.10% accuracy using the DUSE-expanded set at 4% expansion.
  • The active-learning component is load-bearing: the ablation at 7% expansion shows gains of roughly 23 to 24 accuracy points over a one-pass selection without retraining.
  • The framework preserves class balance better than baselines even without explicit balance constraints, which the paper links to the per-class sample counts and t-SNE cluster structure of the augmented set.

Reading between the lines

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

  • Beyond the paper, the same margin score could be combined with a diversity or coverage constraint, for example clustering selected samples in feature space, to avoid selecting many near-duplicate boundary samples; this is a natural next experiment the paper does not run.
  • Beyond the paper, the framework's motivating scenario of a 'related but different' auxiliary domain is untested, since all experiments reuse the same dataset's leftover 99% as the auxiliary pool; a cross-dataset test with overlapping label sets would directly probe that motivation.
  • Beyond the paper, because experiments are restricted to SNR > 10 dB, the low-SNR regime where decision boundaries are much noisier may behave differently; testing DUSE across the full SNR range would show whether margin-based uncertainty still helps.
  • Beyond the paper, the margin score's interpretation as boundary proximity could be tested directly by measuring the distance of selected samples to the true decision boundary in a controlled synthetic signal model, which would separate scoring quality from the active-learning loop's effect.
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

7 major / 6 minor

Summary. The paper proposes DUSE, a data expansion framework for low-resource automatic modulation recognition. Given a small labeled target set D_T and a large labeled auxiliary set D_A, DUSE iteratively trains a classifier on the current target set, scores every auxiliary sample by the margin-based uncertainty u = p*_1 - p*_2, selects the top-K most uncertain samples, moves them into the target set, and repeats. The authors evaluate DUSE against eight coreset-selection baselines on three AMR benchmarks (RML2016.10a, Sig2019-12, RML2018.01a, the latter labeled 'RML2018.10a' in some tables) under class-imbalanced and class-balanced settings, and report cross-architecture generalization of the selected dataset. The abstract claims that DUSE 'consistently outperforms 8 coreset selection baselines in both class-balance and class-imbalance settings,' but the paper's own tables and text contain counterexamples to this claim.

Significance. The idea of selecting informative samples from an already-labeled auxiliary pool via an active-learning-refined margin score is practical and clearly formulated, and the paper gives a formal problem definition that is useful for future work. The computational simplicity of the scorer and the ablation showing a large gain from the active-learning loop are also strengths. However, the headline claim of consistent superiority is not supported by the reported data: DUSE loses to several baselines at the smallest expansion rates on two datasets and at the largest rate on another. The class-balanced experiments cover only one dataset, and the cross-architecture experiments lack any baseline comparison. If these issues are fixed with additional experiments and a more careful claim, the framework would be a plausible practical tool for low-resource AMR; in its current form, the evidence is substantially weaker than advertised.

major comments (7)
  1. [Abstract and Section IV.E, Table I] The abstract's claim that DUSE 'consistently outperforms 8 coreset selection baselines' is contradicted by the paper's own results. In Table I, at 1% expansion on Sig2019-12, DUSE obtains 20.98 ± 0.27 versus Herding's 38.33 ± 12.0 and Margin's 30.69 ± 9.43; at 1% on RML2016.10a, DUSE's 48.68 ± 1.44 trails Forgetting's 51.80 ± 2.42; and at 19% on RML2018.10a, DUSE's 50.45 ± 0.17 trails Forgetting's 55.86 ± 0.38. The text in Section IV.E concedes that 'DUSE does not reach the optimal result under certain expansion rates.' The claimed universal superiority should be revised to a dataset- and budget-dependent statement.
  2. [Section IV.A] The experimental setup does not test the paper's motivating scenario. The introduction argues that the auxiliary dataset comes from 'a related but different field,' but Section IV.A defines D_A as the remaining 99% of the same training set, 'maintaining identical per-class proportions to ensure distributional equivalence.' This means the auxiliary pool has the same label space, same SNR range, and same channel characteristics as the target set. No experiment uses a genuinely different auxiliary domain, so the central claim that DUSE is useful for transfer from a related but different dataset is untested. Either add cross-dataset experiments (e.g., target RML2016.10a with auxiliary Sig2019-12) or reframe the contribution.
  3. [Section IV.E, Table II] The class-balanced experiments are reported only for RML2016.10a; Table II contains no results for Sig2019-12 or RML2018.10a. The abstract and Section IV.E nevertheless claim that DUSE outperforms baselines in 'both class-balance and class-imbalance settings.' The claim is unsupported for class balance beyond a single dataset. Please add class-balanced results for all three datasets or explicitly limit the claim.
  4. [Section IV.E, Table III] The cross-architecture generalization claim is not supported by a comparison. Table III reports accuracies of the DUSE-selected expanded dataset on 1D-CNN, 2D-CNN, AlexNet, SigNet, GRU, and MCLDNN, but it does not compare against expanded datasets produced by any baseline or by random selection. Without such a comparison, the table only shows that the selected set is usable for other architectures, not that DUSE generalizes better than alternatives. Please add cross-architecture evaluations with at least one baseline and random selection.
  5. [Section III.B, Algorithm 2, Eq. (10)] There is an internal inconsistency in the selection direction. Equation (7) defines u = p*_1 - p*_2 and the text states that a smaller u indicates higher uncertainty. Algorithm 2, however, sorts u in descending order and then identifies the 'top-K most uncertain samples,' which would select the largest u, i.e., the least uncertain samples. Equation (10) says 'top-K min u,' which is consistent with selecting small u. The algorithm, equation, and text must be aligned; otherwise the reported DUSE results are not reproducible from the description.
  6. [Section IV.D and Section IV.E] The experimental protocol omits the active-learning round count R and the per-round budget K. Expansion rate r is defined in Eq. (16) as |D*_T - D_T| / |D_A|, but the total number of added samples is R·K, and the paper never reports R or K for the tested rates 1%, 4%, 7%, 9%, and 19%. Without these values, the exact selection procedure cannot be reproduced. Please report R and K, or equivalently the sequence of per-round selection sizes, for each dataset and rate.
  7. [Section IV.A, Tables I and III] The dataset name is inconsistent: the data description in Section IV.A defines RML2018.01a, but Tables I and III and Section IV.E refer to RML2018.10a. If this is a typo, it should be corrected; if a different dataset was actually used, it must be described and cited.
minor comments (6)
  1. [Equation (16)] The denominator D_A is a set, not a number; the formula should read r = |D*_T \ D_T| / |D_A| or similar.
  2. [Equation (3)] The loss term L(f(x_i; theta), y_j) uses index j but the sample is (x_i, y_i); this should be L(f(x_i; theta), y_i).
  3. [Section III.A, Eq. (1)] The notation 'DA ∈ {xA_i, yA_i}' is not a set-builder expression; it should be written as D_A = {(x_i^A, y_i^A)}_{i=1}^{N_A}.
  4. [Section IV.D] The implementation details do not specify the optimizer, the random seed handling, or the justification for restricting to SNR > 10 dB. Since this filter removes a large portion of the data (e.g., 16 of 20 SNR levels in RML2016.10a), its effect on the conclusions should be discussed.
  5. [Section IV.G, Table IV] The 'DUSE (w/o active learning)' row in Table IV is numerically identical to the Margin baseline in Table I at the 7% expansion rate (e.g., 48.77 ± 2.18 on RML2016.10a). This is expected because the one-pass margin scorer is exactly the Margin baseline, but the paper should acknowledge this equivalence explicitly.
  6. [Throughout] There are scattered grammatical and typographical issues, such as 'DUSE is consist of' in Section III and 'for i = 1 to R do' in Algorithm 2. A careful proofreading pass is recommended.

Circularity Check

1 steps flagged · score 1.0 of 10

No significant circularity: DUSE is an empirical active-learning framework whose central claim does not reduce to a fit or to a self-citation chain; the only mild issue is that Eq. (7) renames the standard margin score, which is also one of the baselines.

  1. renaming known result [Section III-A, Eq. (7); Section IV-B, Margin baseline]
    "The uncertainty scoring function for x is then defined as: u = p∗1 − p∗2. A smaller u indicates higher model uncertainty on x. ... Margin [55]. Compute 1−(P(ŷ|x)−maxy≠ŷ P(y|x)), with smaller margins (higher scores) indicating greater confusion between top two classes."

    Eq. (7) defines DUSE's uncertainty score as the gap between the top two class probabilities, which is exactly the Margin baseline score up to the affine transform 1−u; the method section presents this as a newly introduced scoring function, so the scoring component is a renaming of a known baseline. This is a mild reduction by construction, but it does not make the central claim circular: DUSE's reported gains come from the iterative active-learning retraining in Algorithm 2, which the Margin baseline does not use, and all results are evaluated on held-out test splits.

full rationale

There is no fitting of a parameter to the evaluation set and no prediction that reduces to its inputs: the uncertainty score is a fixed formula, the scorer is trained only on the target (and later augmented) data, and accuracy is measured on held-out test splits of the same public datasets. Self-citations (refs. [19]-[21], [42]) appear only in related work and are not load-bearing. The target/auxiliary split (1% vs 99% of the same training set) ensures distributional equivalence and is a limitation of the experiments relative to the 'related but different field' motivation, but it is not circularity. The abstract's universal 'consistently outperforms' claim is contradicted by the paper's own Tables I and II at some expansion rates, but that is a correctness/accuracy concern, not a circular-derivation concern. The only identifiable pattern is renaming the margin score as the 'Uncertainty Scoring Function'; because the active-learning iteration carries the contribution, this does not force the headline result, so the circularity score is minimal.

Assumptions & free parameters 3 free parameters · 2 assumptions · 0 invented entities

No new physical or theoretical entities are introduced. The free parameters are experimental design choices: the SNR cutoff, the unstated round schedule, and the stochastic target initialization. The two domain assumptions are the heuristic link between margin and informativeness, and the distributional-equivalence condition that is never varied in experiments.

free parameters (3)
  • SNR threshold = >10 dB
    All experiments use only samples with SNR > 10 dB (Section IV.A). This hand-chosen cutoff changes the task and is not justified; results may not hold at lower SNR.
  • Number of active-learning rounds R and per-round budget K = Not reported
    Algorithm 2 requires R and K, but the experiments only report the total expansion rate r. The division of the budget across rounds is unspecified and may affect results.
  • Random 1% target subset initialization = Not reported (seed not given)
    The target set is a random 1% class-balanced subset; the choice of seed affects which samples are selected and thus the results. No seeds are reported.
assumptions (2)
  • domain assumption Samples with small margin (p1* - p2*) lie near decision boundaries and are most informative for refining the classifier.
    Section III.A states this without proof; the entire selection strategy rests on this heuristic.
  • domain assumption The auxiliary pool and target set share label space and feature distribution, so selected auxiliary samples improve target classification.
    Section IV.A constructs DA as the remainder of the same training split, ensuring distributional equivalence; the motivating 'related but different' domain setting is not tested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DUSE: A Data Expansion Framework for Low-resource Automatic Modulation Recognition based on Active Learning." pith.science (2026). https://pith.science/paper/23UXQTRK

@misc{pith2026250712011,
  author       = {Pith},
  title        = {Pith review of: DUSE: A Data Expansion Framework for Low-resource Automatic Modulation Recognition based on Active Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/23UXQTRK}},
  note         = {Machine review of arXiv:2507.12011}
}
read the original abstract

Although deep neural networks have made remarkable achievements in the field of automatic modulation recognition (AMR), these models often require a large amount of labeled data for training. However, in many practical scenarios, the available target domain data is scarce and difficult to meet the needs of model training. The most direct way is to collect data manually and perform expert annotation, but the high time and labor costs are unbearable. Another common method is data augmentation. Although it can enrich training samples to a certain extent, it does not introduce new data and therefore cannot fundamentally solve the problem of data scarcity. To address these challenges, we introduce a data expansion framework called Dynamic Uncertainty-driven Sample Expansion (DUSE). Specifically, DUSE uses an uncertainty scoring function to filter out useful samples from relevant AMR datasets and employs an active learning strategy to continuously refine the scorer. Extensive experiments demonstrate that DUSE consistently outperforms 8 coreset selection baselines in both class-balance and class-imbalance settings. Besides, DUSE exhibits strong cross-architecture generalization for unseen models.

Figures

Figures reproduced from arXiv: 2507.12011 by the authors.

Figure 1
Figure 1. Framework Overview. DUSE is consist of two components: (1) an Uncertainty Scoring Function that estimates the informativeness of each sample relative to the target dataset, and (2) an Active Learning Strategy that progressively augments the target dataset using this score. Algorithm 1 Uncertainty Scoring Function Input: A deep learning model f(·; θ), input sample x, number of classes C. Output: Uncertainty score u. … view at source ↗
Figure 2
Figure 2. Visualization of the samples with the highest and lowest uncertainty [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Visualization of the number of samples per class in the expanded dataset obtained by various coreset selection baselines and DUSE (under the setting [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: T-SNE visualization of the expanded dataset obtained by various coreset selection baselines and DUSE (under the setting of class imbalance). [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Continue with ORCID 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

  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...

Reference graph

Works this paper leans on

64 extracted references · 50 canonical work pages · cited by 2 Pith papers

  1. [1]

    Towards automated 3d evaluation of water leakage on a tunnel face via improved gan and self-attention dl model,

    C. Wu, H. Huang, L. Zhang, J. Chen, Y . Tong, and M. Zhou, “Towards automated 3d evaluation of water leakage on a tunnel face via improved gan and self-attention dl model,” Tunnelling and Underground Space Technology, vol. 142, p. 105432, 2023

    work page 2023

  2. [2]

    Feature map distillation of thin nets for low-resolution object recognition,

    Z. Huang, S. Yang, M. Zhou, Z. Li, Z. Gong, and Y . Chen, “Feature map distillation of thin nets for low-resolution object recognition,” IEEE Transactions on Image Processing , vol. 31, pp. 1364–1379, 2022

    work page 2022

  3. [3]

    Fine-grained learning behavior-oriented knowledge distillation for graph neural net- works,

    K. Liu, Z. Huang, C.-D. Wang, B. Gao, and Y . Chen, “Fine-grained learning behavior-oriented knowledge distillation for graph neural net- works,” IEEE Transactions on Neural Networks and Learning Systems , 2024

    work page 2024

  4. [4]

    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

    work page 2016

  5. [5]

    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

    work page 2019

  6. [6]

    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

    work page 2017

  7. [7]

    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

    work page 2017

  8. [8]

    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

    work page 2016

Show all 64 references
  1. [9]

    Minimizing long- term energy consumption in ris-assisted aav-enabled mec network,

    Z. Yao, Q. Zhu, Y . Zhang, H. Huang, and M. Luo, “Minimizing long- term energy consumption in ris-assisted aav-enabled mec network,” IEEE Internet of Things Journal , 2025

    2025

  2. [10]

    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

  3. [11]

    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

  4. [12]

    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

  5. [13]

    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

    2021

  6. [14]

    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

  7. [15]

    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

  8. [16]

    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

  9. [17]

    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

  10. [18]

    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

  11. [19]

    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

  12. [20]

    Fcos: A two-stage recoverable model pruning framework for automatic modulation recognition,

    Y . Lu, T. Ma, Z. Wang, Z. Chen, D. Xu, Y . Lin, Q. Xuan, and G. Gui, “Fcos: A two-stage recoverable model pruning framework for automatic modulation recognition,” arXiv preprint arXiv:2505.21571 , 2025

    2025 arXiv

  13. [21]

    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

  14. [22]

    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

  15. [23]

    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

  16. [24]

    Wisig: A large-scale wifi signal dataset for receiver and channel agnostic rf fingerprinting,

    S. Hanna, S. Karunaratne, and D. Cabric, “Wisig: A large-scale wifi signal dataset for receiver and channel agnostic rf fingerprinting,” IEEE Access, vol. 10, pp. 22 808–22 818, 2022

    2022

  17. [25]

    Trust in 5g open rans through machine learning: Rf fingerprinting on the powder pawr platform,

    G. Reus-Muns, D. Jaisinghani, K. Sankhe, and K. R. Chowdhury, “Trust in 5g open rans through machine learning: Rf fingerprinting on the powder pawr platform,” in GLOBECOM 2020-2020 IEEE Global Communications Conference. IEEE, 2020, pp. 1–6

    2020

  18. [26]

    A simple data augmentation method for automatic modulation recognition via mixing signals,

    W. Shen, D. Xu, X. Xu, Z. Chen, Q. Xuan, W. Wang, Y . Lin, and X. Yang, “A simple data augmentation method for automatic modulation recognition via mixing signals,” in 2024 6th International Conference on Robotics, Intelligent Control and Artificial Intelligence (RICAI) . IEEE...

    2024

  19. [27]

    An efficient data augmentation method for automatic modulation recognition from low- data imbalanced-class regime,

    S. Wei, Z. Sun, Z. Wang, F. Liao, Z. Li, and H. Mi, “An efficient data augmentation method for automatic modulation recognition from low- data imbalanced-class regime,” Applied Sciences, vol. 13, no. 5, p. 3177, 2023

    2023

  20. [28]

    Data augmentation with conditional gan for automatic modulation classification,

    M. Patel, X. Wang, and S. Mao, “Data augmentation with conditional gan for automatic modulation classification,” in Proceedings of the 2nd ACM Workshop on wireless security and machine learning , 2020, pp. 31–36

    2020

  21. [29]

    Spectrum interference- based two-level data augmentation method in deep learning for auto- matic modulation classification,

    Q. Zheng, P. Zhao, Y . Li, H. Wang, and Y . Yang, “Spectrum interference- based two-level data augmentation method in deep learning for auto- matic modulation classification,” Neural Computing and Applications , vol. 33, no. 13, pp. 7723–7745, 2021

    2021

  22. [30]

    Data augmentation aided automatic modulation recognition using diffusion model,

    J. Chen, C. Zhao, X. Huang, and Z. Wu, “Data augmentation aided automatic modulation recognition using diffusion model,” in 2024 IEEE Wireless Communications and Networking Conference (WCNC). IEEE, 2024, pp. 1–6

    2024

  23. [31]

    Diffusion model empowered data augmentation for automatic modulation recognition,

    M. Li, P. Wang, Y . Dong, and Z. Wang, “Diffusion model empowered data augmentation for automatic modulation recognition,” IEEE Wireless Communications Letters, 2025

    2025

  24. [32]

    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

  25. [33]

    Smaller coresets for k-median and k- means clustering,

    S. Har-Peled and A. Kushal, “Smaller coresets for k-median and k- means clustering,” in Proceedings of the twenty-first annual symposium on Computational geometry , 2005, pp. 126–134

    2005

  26. [34]

    A unified framework for approximating and clustering data,

    D. Feldman and M. Langberg, “A unified framework for approximating and clustering data,” in Proceedings of the forty-third annual ACM symposium on Theory of computing , 2011, pp. 569–578

    2011

  27. [35]

    Core vector machines: Fast svm training on very large data sets

    I. W. Tsang, J. T. Kwok, N. Cristianini et al. , “Core vector machines: Fast svm training on very large data sets.” Journal of machine Learning research, vol. 6, no. 4, 2005

    2005

  28. [36]

    Coresets for scalable bayesian logistic regression,

    J. Huggins, T. Campbell, and T. Broderick, “Coresets for scalable bayesian logistic regression,” Advances in neural information processing systems, vol. 29, 2016

    2016

  29. [37]

    Training gaussian mixture models at scale via coresets,

    M. Lucic, M. Faulkner, A. Krause, and D. Feldman, “Training gaussian mixture models at scale via coresets,” Journal of Machine Learning Research, vol. 18, no. 160, pp. 1–25, 2018

    2018

  30. [38]

    Scalable training of mix- ture models via coresets,

    D. Feldman, M. Faulkner, and A. Krause, “Scalable training of mix- ture models via coresets,” Advances in neural information processing systems, vol. 24, 2011

    2011

  31. [39]

    Efficient coreset selection with cluster-based methods,

    C. Chai, J. Wang, N. Tang, Y . Yuan, J. Liu, Y . Deng, and G. Wang, “Efficient coreset selection with cluster-based methods,” in Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2023, pp. 167–178

    2023

  32. [40]

    Deepcore: A comprehensive library for coreset selection in deep learning,

    C. Guo, B. Zhao, and Y . Bai, “Deepcore: A comprehensive library for coreset selection in deep learning,” in International Conference on Database and Expert Systems Applications . Springer, 2022, pp. 181– 195. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 10

    2022

  33. [41]

    A coreset selection of coreset selection literature: Introduction and recent advances,

    B. B. Moser, A. S. Shanbhag, S. Frolov, F. Raue, J. Folz, and A. Dengel, “A coreset selection of coreset selection literature: Introduction and recent advances,” arXiv preprint arXiv:2505.17799 , 2025

    2025

  34. [42]

    Rk-core: An established methodology for exploring the hierarchical structure within datasets,

    Y . Lu, Y . Huang, J. Nie, Z. Chen, and Q. Xuan, “Rk-core: An established methodology for exploring the hierarchical structure within datasets,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2024, pp. 3150–3154

    2024

  35. [43]

    Coreset selection for object detection,

    H. Lee, S. Kim, J. Lee, J. Yoo, and N. Kwak, “Coreset selection for object detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 7682–7691

    2024

  36. [44]

    Coreset selection via reducible loss in continual learning,

    R. Tong, Y . Liu, J. Q. Shi, and D. Gong, “Coreset selection via reducible loss in continual learning,” in The Thirteenth International Conference on Learning Representations , 2025

    2025

  37. [45]

    Fedcs: Coreset selection for federated learning,

    C. Hao, W. Xie, D. Li, H. Qin, H. Ye, L. Fang, and Y . Li, “Fedcs: Coreset selection for federated learning,” in Proceedings of the Computer Vision and Pattern Recognition Conference , 2025, pp. 15 434–15 443

    2025

  38. [46]

    Goodcore: Data-effective and data-efficient machine learning through coreset selection over incomplete data,

    C. Chai, J. Liu, N. Tang, J. Fan, D. Miao, J. Wang, Y . Luo, and G. Li, “Goodcore: Data-effective and data-efficient machine learning through coreset selection over incomplete data,” Proceedings of the ACM on Management of Data , vol. 1, no. 2, pp. 1–27, 2023

    2023

  39. [47]

    Active learning literature survey,

    B. Settles, “Active learning literature survey,” 2009

    2009

  40. [48]

    Learning from human educational wisdom: A student-centered knowledge distillation method,

    S. Yang, J. Yang, M. Zhou, Z. Huang, W.-S. Zheng, X. Yang, and J. Ren, “Learning from human educational wisdom: A student-centered knowledge distillation method,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 46, no. 6, pp. 4188–4205, 2024

    2024

  41. [49]

    A cost-sensitive active learning algorithm: toward imbalanced time series forecasting,

    J. Zhang and Q. Dai, “A cost-sensitive active learning algorithm: toward imbalanced time series forecasting,” Neural Computing and Applications, vol. 34, no. 9, pp. 6953–6972, 2022

    2022

  42. [50]

    Cost sensitive active learning using bidirectional gated recurrent neural networks for imbalanced fault diagnosis,

    P. Peng, Z. Wenjia, Z. Yi, X. Yanyan, W. Hongwei, and Z. Heming, “Cost sensitive active learning using bidirectional gated recurrent neural networks for imbalanced fault diagnosis,” Neurocomputing, vol. 407, pp. 232–245, 2020

    2020

  43. [51]

    Task-aware variational adversarial active learning,

    K. Kim, D. Park, K. I. Kim, and S. Y . Chun, “Task-aware variational adversarial active learning,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 8166–8175

    2021

  44. [52]

    Active transfer learning network: A unified deep joint spectral–spatial feature learning model for hyperspectral image classification,

    C. Deng, Y . Xue, X. Liu, C. Li, and D. Tao, “Active transfer learning network: A unified deep joint spectral–spatial feature learning model for hyperspectral image classification,” IEEE Transactions on Geoscience and Remote Sensing , vol. 57, no. 3, pp. 1741–1754, 2018

    2018

  45. [53]

    Modulation and signal class labelling with active learning and classification using ma- chine learning,

    B. Bhargava, A. Deshmukh, and A. Narasimhadhan, “Modulation and signal class labelling with active learning and classification using ma- chine learning,” in 2022 IEEE International Conference on Electronics, Computing and Communication Technologies (CONECCT) . IEEE, 2022, pp. 1–5

    2022

  46. [54]

    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

  47. [55]

    Selection via proxy: Efficient data selection for deep learning,

    C. Coleman, C. Yeh, S. Mussmann, B. Mirzasoleiman, P. Bailis, P. Liang, J. Leskovec, and M. Zaharia, “Selection via proxy: Efficient data selection for deep learning,” arXiv preprint arXiv:1906.11829 , 2019

    1906 arXiv

  48. [56]

    Adversarial active learning for deep networks: a margin based approach,

    M. Ducoffe and F. Precioso, “Adversarial active learning for deep networks: a margin based approach,” arXiv preprint arXiv:1802.09841 , 2018

    2018 arXiv

  49. [57]

    An empirical study of example forgetting during deep neural network learning,

    M. Toneva, A. Sordoni, R. T. d. Combes, A. Trischler, Y . Bengio, and G. J. Gordon, “An empirical study of example forgetting during deep neural network learning,” arXiv preprint arXiv:1812.05159 , 2018

    2018 arXiv

  50. [58]

    Glister: Generalization based data subset selection for efficient and robust learning,

    K. Killamsetty, D. Sivasubramanian, G. Ramakrishnan, and R. Iyer, “Glister: Generalization based data subset selection for efficient and robust learning,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 35, no. 9, 2021, pp. 8110–8118

    2021

  51. [59]

    Deep learning on a data diet: Finding important examples early in training,

    M. Paul, S. Ganguli, and G. K. Dziugaite, “Deep learning on a data diet: Finding important examples early in training,” Advances in neural information processing systems , vol. 34, pp. 20 596–20 607, 2021

    2021

  52. [60]

    Herding dynamical weights to learn,

    M. Welling, “Herding dynamical weights to learn,” in Proceedings of the 26th annual international conference on machine learning , 2009, pp. 1121–1128

    2009

  53. [61]

    Super-samples from kernel herding,

    Y . Chen, M. Welling, and A. Smola, “Super-samples from kernel herding,” arXiv preprint arXiv:1203.3472 , 2012

    2012 arXiv

  54. [62]

    Imagenet classification with deep convolutional neural networks,

    A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” Advances in neural informa- tion processing systems , vol. 25, 2012

    2012

  55. [63]

    A spatiotemporal multi-channel learning framework for automatic modulation recognition,

    J. Xu, C. Luo, G. Parr, and Y . Luo, “A spatiotemporal multi-channel learning framework for automatic modulation recognition,” IEEE Wire- less Communications Letters , vol. 9, no. 10, pp. 1629–1632, 2020

    2020

  56. [64]

    Stochastic neighbor embedding,

    G. E. Hinton and S. Roweis, “Stochastic neighbor embedding,” Advances in neural information processing systems , vol. 15, 2002. Yao Lu received his B.S. degree from Zhejiang University of Technology and is currently pursuing a Ph.D. in control science and engineering at Zhejia...

    2002

Pith tools

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