Pith. sign in

REVIEW 4 major objections 5 minor 25 references

This paper claims that an uncertainty-driven hybrid network—a fast 2D CNN plus a BiLSTM triggered by an MC Dropout uncertainty score—can recognize 14 RF modulation classes at 92.6% accuracy while keeping the fast path at 0.138 ms per sample

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review

2026-08-05 00:10 UTC pith:KGK4WUXS

load-bearing objection Useful engineering extension with a genuinely new uncertainty-routing rule, but the headline accuracy and real-time claims are tied to a validation-tuned threshold the paper never sensitivity-tests. the 4 major comments →

arxiv 2608.00796 v1 pith:KGK4WUXS submitted 2026-08-01 eess.SP cs.LG

An Uncertainty-Driven Hybrid Deep Learning Approach for Broad-Coverage RF Modulation Recognition

Nurettin Safak , Durdu Can Yerdeyatar , Muhammet Sefa Demirel , Alperen Marasli , Taha Eren Atmaca , Ozgun Ersoy This is my paper
classification eess.SP cs.LG
keywords RF signal recognitionautomatic modulation classificationhybrid deep learninguncertainty estimationBayesian learningMC Dropout2D CNNBiLSTM
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Automatic modulation recognition—deciding what type of radio signal is being received—is hard at low signal-to-noise ratio and across a wide class of modulations. The paper proposes a hybrid system that first classifies quickly with a small 2D CNN on a compact spectrogram, then uses an MC Dropout-based uncertainty score to decide whether that fast answer can be trusted; for low-confidence cases it routes the signal to a slower BiLSTM that reads the temporal IQ structure. On a 14-class simulation dataset the fast path alone reaches 83.3% accuracy at 0.138 ms per sample, and the full hybrid reaches 92.6% accuracy on an end-to-end validation set, with the uncertainty score detecting CNN misclassifications at AUROC 0.78. The paper's point is that a fast path plus uncertainty-triggered slow path can get near-sequential accuracy at near-real-time latency, and that compact spectral features are structurally insufficient for FSK-order disambiguation such as 4-FSK.

Core claim

The central claim is that an uncertainty-driven hybrid architecture can achieve broad-coverage RF modulation recognition without paying BiLSTM latency on every sample. The decision flow uses FFT-extracted parameters and a 32x32 STFT spectrogram as input; the 2D CNN produces a primary label in 0.138 ms, an MC Dropout Bayesian MLP computes a per-sample uncertainty score as the maximum over classes of the standard deviation of thirty stochastic softmax passes, and if that score exceeds a threshold the sample is handed to a two-layer BiLSTM that processes the first 1024 IQ samples. If the score is far above threshold, the sample is rejected as out-of-distribution. In end-to-end validation the au

What carries the argument

The load-bearing mechanism is the uncertainty score computed from thirty MC Dropout passes of a Bayesian MLP, paired with a threshold that routes samples between the fast 2D CNN and the slower BiLSTM. The score does the work of predicting when the CNN is likely wrong, so the system keeps latency low for most samples and spends 15.2 ms only on those the fast path cannot be trusted on. The compact 32x32 log-power STFT spectrogram is the input representation that makes the CNN fast; the paper shows that this representation cannot separate 4-FSK from 2-FSK, which is why temporal modeling in the BiLSTM is needed.

Load-bearing premise

The reported 92.6% hybrid accuracy depends on a single routing threshold (0.05) chosen on the validation set, and the paper assumes that threshold will keep working for unseen SNR levels and channel conditions—it never tests a fixed threshold across SNR tiers, and the same threshold is stated as 0.15 earlier in the paper.

What would settle it

Run the fixed routing threshold on held-out SNR blocks and on Rayleigh/Rician test sets without re-tuning; if the hybrid accuracy or the AUROC for misclassification detection collapses to the fast path's level, the claim that uncertainty routing generalizes is false. Concretely, repeat the Section IV-E end-to-end validation with the threshold fixed across all seven SNR levels, and with a threshold chosen on AWGN then evaluated on fading channels.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • A fixed-budget recognition system can use the uncertainty score to decide per sample whether to spend 0.138 ms or 15.2 ms, so average latency depends on the routed fraction (38% in the validation set).
  • Compact spectral features are sufficient for many modulation classes but not for FSK order; any fast classifier using this representation will need temporal or phase-preserving features, or an auxiliary resolver for FSK and 4-FSK.
  • MC Dropout with thirty passes and a dropout rate of 0.3 provides a usable misclassification predictor (AUROC 0.78) at 0.7 ms, well below the BiLSTM's 15.2 ms, making it a practical routing signal.
  • Mixed-channel training on AWGN, Rayleigh, and Rician channels improves fading-channel accuracy by 3.6 to 5.6 points without hurting AWGN accuracy, so the hybrid can be adapted to fading environments cheaply.
  • Evaluation sets that omit 4-FSK overstate performance for compact-spectral methods; broad-coverage benchmarks should include FSK orders and spread-spectrum classes.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper, a natural testable extension is to choose the routing threshold separately for each SNR tier or make it a function of estimated SNR; if the reported 92.6% hybrid gain persists under per-tier thresholds, the routing idea is more robust than a single global threshold.
  • The AUROC 0.78 for misclassification prediction is measured on a closed 14-class set; an open-set deployment with unknown modulations would need the out-of-distribution rejection branch validated separately, since the paper leaves that to future work.
  • Because the fast path uses a compact spectrogram and the slow path uses raw IQ, the slow path could be replaced by any temporally-aware sequence model and the routing mechanism would still apply; the paper only demonstrates the idea with a BiLSTM.
  • A comparison at a fixed average latency budget, rather than at each architecture's natural latency, would clarify whether the hybrid beats a single mid-size model; the paper compares accuracy at fixed architecture latency, not accuracy at a fixed average latency.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. This paper proposes a hybrid automatic modulation classification (AMC) system for 14 modulation classes, combining a fast 2D CNN primary path on 32×32 log-power STFT spectrograms with an MC-Dropout Bayesian MLP uncertainty estimator and a BiLSTM secondary path that is activated when the uncertainty score exceeds a threshold. The 2D CNN is reported at 83.3±0.7% accuracy and 0.138 ms per sample. On a separate 29,400-sample end-to-end validation set, the uncertainty score predicts CNN misclassifications with AUROC 0.78, and the hybrid system reaches 92.6% accuracy (macro-F1 92.3%) with 38.0% of samples routed to the BiLSTM, using a macro-F1-selected threshold τ=0.05. The paper also reports SNR-dependent comparisons against rule-based, SVM, XGBoost, Bayesian MLP, and BiLSTM baselines, ablations of input resolution, FFT length, number of MC samples, and dropout rate, plus channel robustness and channel-adaptive training results under Rayleigh and Rician fading.

Significance. The core idea of using predictive uncertainty to route only hard examples to an expensive temporal model is practically motivated, and the results are useful if properly validated. The manuscript has clear strengths: controlled signal generation with per-sample random seeds, a relatively broad 14-class set including 4-FSK, DSSS, and FHSS, direct comparisons in Figures 4–5 and Table III, and a threshold-independent AUROC (0.78) that supports the utility of the uncertainty score. The 4-FSK recall result (3% for the spectral MLP) is an honest and informative negative finding. However, the central quantitative claims are currently conditional on a validation-tuned threshold and on an unsupported real-time assertion, and the OOD-rejection branch is not evaluated. These issues need to be resolved before the contribution is fully established.

major comments (4)
  1. [§III-C, §IV-E, Table IV] The routing threshold is inconsistent and the headline results are validation-tuned. Section III-C states τ=0.15 is 'empirically selected', while Section IV-E and Table IV report the end-to-end results with τ=0.05, selected by macro-F1 on the validation split. The reported 92.6% hybrid accuracy and 38.0% routing fraction are therefore conditional on this threshold choice, and no τ-sensitivity analysis, per-SNR routing rates, or per-channel routing rates are given. As a result, the central numerical claims are not shown to generalize to a fixed operational threshold. Please report hybrid accuracy and routing fraction as functions of τ, and validate the selected threshold on a truly held-out test split.
  2. [Abstract, §III-A, Table III, §IV-E] The real-time/low-latency claim is not supported by the measurements. The 0.138 ms figure is only the 2D CNN forward pass; every sample also requires FFT/STFT preprocessing and a 0.7 ms Bayesian MLP pass, and 38% of samples are routed to the 15.2 ms BiLSTM. Even ignoring preprocessing, the mean inference cost is approximately 0.62×(0.138+0.7) + 0.38×(0.138+0.7+15.2) ≈ 6.6 ms, with a 15.2 ms tail. No latency budget, throughput measurement, or percentile latency is provided. The real-time claim should be supported by an end-to-end latency measurement on the actual hardware and an explicit application-defined latency budget.
  3. [Algorithm 1, Table II, §V] Algorithm 1 (line 11) returns OUT-OF-DISTRIBUTION when U(x)≫τ, and Table II credits the hybrid system with OOD detection, but no OOD experiment is reported. The 14-class closed set contains no out-of-distribution signals, no rejection rate or false-positive rate is given, and the condition '≫' is not specified. The conclusion acknowledges that comprehensive OOD validation is future work; this should be reconciled with the contribution claims, either by adding an open-set experiment (e.g., unknown modulations or noise-only inputs) or by removing the OOD claim from the architecture comparison.
  4. [Table IV, §IV-E, §V] The evaluation protocol for the 92.6% headline accuracy is ambiguous. Table IV reports a 70/15/15 split and states the threshold is macro-F1-selected, but does not state whether the reported 92.6%/92.3% numbers are computed on the validation or test portion. Since threshold selection is performed on the validation split, reporting accuracy on that same split would be optimistic. Moreover, this end-to-end result comes from a different dataset than the 83.3±0.7% 2D CNN result, so the implied comparison in the conclusion ('raises hybrid accuracy to 92.6%') is not apples-to-apples. Please specify which split is used for each headline number, and ideally report test-set performance with confidence intervals across multiple runs.
minor comments (5)
  1. [§III-B, Fig. 2] The 2D CNN parameter count is given as ≈289,745 in the text but ~360k in Fig. 2. Please reconcile these numbers.
  2. [§IV-E] The text reports 'AUROC 0.78 at τ=0.05'; AUROC is threshold-independent, so the parenthetical 'at τ=0.05' is misleading and should be removed.
  3. [§IV-B] The 199,920-sample companion benchmark is described as 'reused rather than re-run'. Please clarify whether this dataset is independent from the Section II-A dataset and describe its split and random-seed handling for reproducibility.
  4. [§IV-C] The 4-FSK recall values (3%, 71%, near-perfect for BiLSTM) are stated without a confusion matrix or per-class table. Adding a per-class breakdown would strengthen the FSK-disambiguation argument and the motivation for routing.
  5. [Table II, §IV-C] The qualitative ratings 'Low/Med./High' for 4-FSK recall in Table II are not backed by a numeric table. Consider replacing these with the quantitative recall values reported in Section IV-C.

Circularity Check

0 steps flagged

No significant circularity: all headline metrics are measured on held-out test data; threshold selection is standard tuning.

full rationale

The paper's central claims are empirical measurements on held-out test splits. The only fitted parameter in the end-to-end validation is the routing threshold τ, which is selected on a validation split (Table IV: 'Threshold τ 0.05 (macro-F1-selected)'); the reported 92.6% hybrid accuracy and 38.0% routed fraction are then evaluated on the test split, so the headline numbers are not forced by construction. The uncertainty score's AUROC (0.78) is threshold-independent and measures the correlation between the Bayesian MLP's MC-Dropout standard deviation and CNN misclassifications; the MLP is trained on class labels, not on CNN errors, so the measurement is not definitionally circular. There are no self-citations, no imported uniqueness theorems, and no ansatz smuggled in via citation: MC Dropout is adopted from external prior work [6], [15]. The inconsistency between τ=0.15 (Section III-C) and τ=0.05 (Section IV-E) is a robustness/threshold-selection concern, not a circularity, because no reported metric is derived from the threshold itself. The paper is self-contained against external benchmarks and its results are measured rather than derived.

Axiom & Free-Parameter Ledger

1 free parameters · 3 axioms · 0 invented entities

The system introduces no new physical entities or conserved quantities; all components are standard (CNN, MLP, LSTM, FFT, STFT). The main tuning knob is the uncertainty threshold τ, which is fitted to validation data.

free parameters (1)
  • Uncertainty threshold τ = 0.15 (Section III-C), 0.05 (Section IV-E)
    Empirically selected on validation data to balance routing and accuracy; the two values appear without reconciliation.
axioms (3)
  • domain assumption AWGN simulation channel with fs=200 kHz, fsym=10 kHz, 4096-sample signals, and specified DSSS/FHSS parameters
    Section II-A defines the entire evaluation environment; real channels are not considered in the main experiments.
  • domain assumption MC Dropout with T=30 stochastic passes approximates the Bayesian posterior predictive uncertainty
    Borrowed from Gal and Ghahramani [6]; the paper uses it as the basis for the routing decision.
  • domain assumption The 14 modulation classes are separable in the 32x32 log-power spectrogram plus auxiliary FFT features
    The paper relies on this for the primary CNN path; ablations show some classes (4-FSK) are not separable.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of An Uncertainty-Driven Hybrid Deep Learning Approach for Broad-Coverage RF Modulation Recognition." pith.science (2026). https://pith.science/paper/KGK4WUXS

@misc{pith2026260800796,
  author       = {Pith},
  title        = {Pith review of: An Uncertainty-Driven Hybrid Deep Learning Approach for Broad-Coverage RF Modulation Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KGK4WUXS}},
  note         = {Machine review of arXiv:2608.00796}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Automatic RF modulation recognition is of critical importance in spectrum monitoring, electronic warfare, and cognitive radio applications, where low signal-to-noise ratio (SNR) conditions and the growing diversity of modulation schemes limit the performance of existing methods. This paper proposes an uncertainty-driven hybrid deep learning architecture for recognizing RF signals over a broad modulation space. The proposed approach carries out a multi-stage classification process by combining spectral information obtained through low-cost FFT-based preprocessing with time-frequency features extracted from short-time Fourier transform (STFT) spectrograms. The architecture comprises a 2D convolutional neural network (2D CNN)-based path for fast, low-latency primary classification, MC Dropout-supported Bayesian uncertainty estimation for assessing classification reliability, and a BiLSTM-based secondary decision mechanism activated under high-uncertainty conditions. The proposed system is evaluated in a controlled simulation environment spanning different SNR levels and modulation classes. Experimental results show that the primary 2D CNN path achieves $83.3\pm0.7\%$ accuracy with an inference time of only 0.138 ms per sample, providing superior performance compared with traditional rule-based and classical machine-learning approaches. Furthermore, the obtained findings reveal the limitations of compact spectral feature representations and classifiers lacking temporal modeling, particularly in disambiguating FSK-based modulations. The uncertainty estimation module offers promising results for detecting low-confidence decisions, and the proposed approach demonstrates the potential of a low-latency and scalable solution for real-time RF modulation recognition.

Figures

Figures reproduced from arXiv: 2608.00796 by Alperen Marasli, Durdu Can Yerdeyatar, Muhammet Sefa Demirel, Nurettin Safak, Ozgun Ersoy, Taha Eren Atmaca.

Figure 1
Figure 1. Figure 1: Hybrid decision flow: FFT extraction and STFT, 2D CNN (0.138 ms), [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 4
Figure 4. Figure 4: Model accuracy comparison by SNR (14 classes, [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Model accuracy at +20 dB (same benchmark as [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: MC Dropout ablation. (A) MC samples T: T = 30 vs. T = 50 differ by <0.2 points. (B) Dropout rate p: best at p = 0.3. (C) With/without dropout: +0.3 points and Bayesian uncertainty capability. TABLE V CHANNEL-ADAPTIVE TRAINING COMPARISON (+20 DB TEST ACCURACY) Training Channel AWGN Rayleigh Rician AWGN only 87.4% 80.0% 82.9% Rayleigh only 85.6% 84.6% 85.7% Mixed (3 channels) 88.2% 85.6% 86.5% (AWGN+Rayleigh… view at source ↗
Figure 7
Figure 7. Figure 7: Channel model comparison: AWGN, Rayleigh, Rician ( [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

25 extracted references · 17 canonical work pages · 3 internal anchors

  1. [1]

    Survey of automatic modulation classification techniques: Classical approaches and new trends,

    O. A. Dobre, A. Abdi, Y . Bar-Ness, and W. Su, “Survey of automatic modulation classification techniques: Classical approaches and new trends,”IET Commun., vol. 1, no. 2, pp. 137–156, 2007. doi: 10.1049/iet- com:20050176

    doi:10.1049/iet- 2007

  2. [2]

    An introduction to deep learning for the physical layer,

    T. J. O’Shea and J. Hoydis, “An introduction to deep learning for the physical layer,”IEEE Trans. Cogn. Commun. Netw., vol. 3, no. 4, pp. 563–575, 2017. doi: 10.1109/TCCN.2017.2758370

    work page arXiv 2017

  3. [3]

    Deep architectures for modulation recog- nition,

    N. E. West and T. O’Shea, “Deep architectures for modulation recog- nition,” inProc. IEEE DySPAN, pp. 1–6, 2017. doi: 10.1109/DyS- PAN.2017.7920754

    work page arXiv 2017

  4. [4]

    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 J. Sel. Topics Signal Process., vol. 12, no. 1, pp. 168–179, 2018. doi: 10.1109/JSTSP.2018.2797022

    work page arXiv 2018

  5. [5]

    Deep learning models for wireless signal classification with distributed low- cost spectrum sensors,

    S. Rajendran, W. Meert, D. Giustiniano, V . Lenders, and S. Pollin, “Deep learning models for wireless signal classification with distributed low- cost spectrum sensors,”IEEE Trans. Cogn. Commun. Netw., vol. 4, no. 3, pp. 433–445, 2018. doi: 10.1109/TCCN.2018.2835460

    work page arXiv 2018

  6. [6]

    Dropout as a Bayesian approximation: Representing model uncertainty in deep learning,

    Y . Gal and Z. Ghahramani, “Dropout as a Bayesian approximation: Representing model uncertainty in deep learning,” inProc. ICML, pp. 1050–1059, 2016. arXiv:1506.02142

    Pith/arXiv arXiv 2016

  7. [7]

    Deep neural network architectures for modulation classification,

    X. Liu, D. Yang, and A. El Gamal, “Deep neural network architectures for modulation classification,” inProc. Asilomar Conf. Signals, Syst. Comput., pp. 915–919, 2017. doi: 10.1109/ACSSC.2017.8335483

    work page arXiv 2017

  8. [8]

    Spectrum analysis and convolutional neural network for automatic modulation recognition,

    Y . Zeng, M. Zhang, F. Han, Y . Gong, and J. Zhang, “Spectrum analysis and convolutional neural network for automatic modulation recognition,” IEEE Wireless Commun. Lett., vol. 8, no. 3, pp. 929–932, 2019. doi: 10.1109/LWC.2019.2900247

    work page arXiv 2019

  9. [9]

    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 Wireless Commun. Lett., vol. 9, no. 10, pp. 1629–1632, 2020. doi: 10.1109/LWC.2020.2999453

    work page arXiv 2020

  10. [10]

    A survey of modulation classification using deep learning: Signal representation and data preprocessing,

    S. Peng, S. Sun, and Y .-D. Yao, “A survey of modulation classification using deep learning: Signal representation and data preprocessing,”IEEE Trans. Neural Netw. Learn. Syst., vol. 33, no. 12, pp. 7020–7038, 2022. doi: 10.1109/TNNLS.2021.3085433

    work page arXiv 2022

  11. [11]

    Sequential convolu- tional recurrent neural networks for fast automatic modulation classifi- cation,

    K. Liao, Y . Zhao, J. Gu, Y . Zhang, and Y . Zhong, “Sequential convolu- tional recurrent neural networks for fast automatic modulation classifi- cation,”IEEE Access, vol. 9, pp. 27182–27188, 2021. doi: 10.1109/AC- CESS.2021.3053427

    arXiv 2021

  12. [12]

    Deep learn- ing based automatic modulation recognition: Models, datasets, and challenges,

    F. Zhang, C. Luo, J. Xu, Y . Luo, and F.-C. Zheng, “Deep learn- ing based automatic modulation recognition: Models, datasets, and challenges,”Digit. Signal Process., vol. 129, art. 103650, 2022. doi: 10.1016/j.dsp.2022.103650

    work page arXiv 2022

  13. [13]

    Automatic modu- lation classification based on CNN-Transformer graph neural network,

    D. Wang, M. Lin, X. Zhang, Y . Huang, and Y . Zhu, “Automatic modu- lation classification based on CNN-Transformer graph neural network,” Sensors, vol. 23, no. 16, art. 7281, 2023. doi: 10.3390/s23167281

    work page doi:10.3390/s23167281 2023

  14. [14]

    LightAMC: Lightweight automatic modulation classification via deep learning and compressive sensing,

    Y . Wang, J. Yang, M. Liu, and G. Gui, “LightAMC: Lightweight automatic modulation classification via deep learning and compressive sensing,”IEEE Trans. Veh. Technol., vol. 69, no. 3, pp. 3491–3495,

    work page

  15. [15]

    What uncertainties do we need in Bayesian deep learning for computer vision?,

    A. Kendall and Y . Gal, “What uncertainties do we need in Bayesian deep learning for computer vision?,” inAdv. Neural Inf. Process. Syst. (NeurIPS), vol. 30, pp. 5574–5584, 2017. arXiv:1703.04977

    Pith/arXiv arXiv 2017

  16. [16]

    Simple and scal- able predictive uncertainty estimation using deep ensembles,

    B. Lakshminarayanan, A. Pritzel, and C. Blundell, “Simple and scal- able predictive uncertainty estimation using deep ensembles,” inAdv. Neural Inf. Process. Syst. (NeurIPS), vol. 30, pp. 6402–6413, 2017. arXiv:1612.01474

    Pith/arXiv arXiv 2017

  17. [17]

    An uncertainty quantification frame- work for deep learning-based automatic modulation classification,

    H. Yang and R. Sahay, “An uncertainty quantification frame- work for deep learning-based automatic modulation classification,” arXiv:2503.04142, 2025

    work page arXiv 2025

  18. [18]

    Open Set Wireless Signal Classification: Augmenting Deep Learning with Expert Feature Classifiers

    S. R. Shebert, B. H. Kirk, and R. M. Buehrer, “Open set wireless signal classification: Augmenting deep learning with expert feature classifiers,” arXiv:2302.03749, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  19. [19]

    Class Information Guided Reconstruction for Automatic Modulation Open-Set Recognition

    Z. Zhang, M. Zhu, J. Liu, Y . Li, and S. Wang, “Class information guided reconstruction for automatic modulation open-set recognition,” arXiv:2312.13023, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  20. [20]

    A baseline for detecting misclassified and out-of-distribution examples in neural networks,

    D. Hendrycks and K. Gimpel, “A baseline for detecting misclassified and out-of-distribution examples in neural networks,” inProc. Int. Conf. Learn. Represent. (ICLR), 2017. arXiv:1610.02136

    Pith/arXiv arXiv 2017

  21. [21]

    Enhancing automatic modulation recognition through robust global feature extraction,

    Y . Qu, Z. Lu, R. Zeng, J. Wang, and J. Wang, “Enhancing automatic modulation recognition through robust global feature extraction,” sub- mitted toIEEE Trans. Veh. Technol., 2024. arXiv:2401.01056

    Pith/arXiv arXiv 2024

  22. [22]

    MAMCA – optimal on accuracy and efficiency for automatic modulation classification with extended signal length,

    Y . Zhang, Z. Zhou, Y . Cao, G. Li, and X. Li, “MAMCA – optimal on accuracy and efficiency for automatic modulation classification with extended signal length,”arXiv:2405.11263, 2024

    Pith/arXiv arXiv 2024

  23. [23]

    AI/ML-Based Automatic Modulation Recognition: Recent Trends and Future Possibilities

    E. Jafarigol, B. Alaghband, A. Gilanpour, S. Hosseinipoor, and M. Mir- mozafari, “AI/ML-based automatic modulation recognition: Recent trends and future possibilities,”arXiv:2502.05315, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  24. [24]

    CNN-LSTM hybrid architecture for over-the-air automatic modulation classification using SDR,

    D. Padhya, K. Acharya, B. K. Dahal, and D. B. Kshatri, “CNN-LSTM hybrid architecture for over-the-air automatic modulation classification using SDR,” accepted inJ. Innov. Eng. Educ., 2025. arXiv:2511.21040

    work page arXiv 2025

  25. [2020]

    doi: 10.1109/TVT.2020.2971001

    work page arXiv 2020

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.