Pith. sign in

REVIEW 4 major objections 4 minor 61 references

Composite modulation recognition can go zero-shot: logarithm turns multiplicative coupling into addition, and a learned affine transform disentangles the layers, recognizing unseen inner-outer pairs at over 93% accuracy.

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 · deepseek-v4-flash

2026-08-02 05:06 UTC pith:LEV7CYKZ

load-bearing objection The system is a legitimate engineering combo and may work for easy holdouts, but the paper's own per-holdout results contradict the 93% headline, and the logged complex-signal theory is not established. the 4 major comments →

arxiv 2607.13463 v1 pith:LEV7CYKZ submitted 2026-07-15 eess.SP

Compositional Zero-Shot Recognition based on Tangent Space Disentanglement for Composite Modulation Signals

Yurui Zhao , Xiang Wang , Zhitao Huang , Baoguo Li This is my paper
classification eess.SP
keywords composite modulation recognitionzero-shot learningtangent spacelogarithmic mappingspatial transformer networkISACsignal disentanglementsemantic space
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.

The paper tries to establish that zero-shot recognition of composite modulation (CM) signals — waveforms formed by multiplying an inner communication modulation by an outer sensing modulation — is achievable without ever training on the target combination. The crucial idea is that the pointwise product s_CM = s_in ⊙ s_out becomes a sum after a logarithmic projection, so the two entangled layers can be separated by learnable affine transformations and matched against per-layer prototypes. If correct, this turns the combinatorial growth of modulation pairs in ISAC systems into a scalable recognition problem: any combination of individually seen modulations can be recognized at test time. The authors report over 93% zero-shot accuracy, outperforming unified-semantic and multi-task baselines, with robustness maintained under combined channel fading and hardware impairments down to 4 dB SNR.

Core claim

On the paper's own terms, the central discovery is that the multiplicative coupling of CM layers is linearized by the homomorphism log : (R+, ×) → (R, +). The paper proves that an input-dependent affine transformation suffices to isolate the CM component from the additive tangent-space representation (Proposition 1), then instantiates this as TSDN, a dual-branch network where a spatial transformer learns the per-layer disentangling transform and a multi-objective loss (cross-entropy, center loss, and a feature orthogonality penalty) shapes the semantic space. The empirical claim is over 93% zero-shot accuracy on unseen combinations, generalization to high-order modulations like 128QAM-8FSK,

What carries the argument

The load-bearing object is the element-wise logarithmic map T(·) = log(·), which is claimed to convert the multiplicative coupling s_CM = s_in ⊙ s_out into the additive decomposition x = x_CM + x_Δ + ñ (Eqs. 21–22) by projecting the signal manifold onto its tangent space at the identity. The second essential piece is the input-dependent affine transformation Θ(x) = [x 1]Θ (learned by a spatial transformer network), which Proposition 1 shows can perfectly isolate x_CM; the third is the factored semantic space of per-layer prototypes for inner and outer modulation types, enabling compositional matching.

Load-bearing premise

The claim stands on the assumption that the element-wise logarithm, applied to the complex received signal, truly converts the multiplicative coupling into exact addition — a property that holds for strictly positive real values, but the paper does not establish it for complex baseband waveforms, where the logarithm is multi-valued modulo 2πi and undefined at zero.

What would settle it

Compute the residual of the additive decomposition x = x_CM + x_Δ + ñ on the actual complex CM signals at a few sample instants: if the imaginary part of log(s_rx) does not match log-magnitude plus a consistent branch choice, or if the residual is comparable in size to the components, the core linearization is falsified. Experimentally, re-run TSDN with the logarithmic mapping replaced by a mathematically well-defined choice (e.g., log of magnitude, or complex log with a fixed branch cut); if zero-shot accuracy does not drop, the claimed homomorphism is not what carries the performance.

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

If this is right

  • If the claim holds, ISAC receivers can recognize composite modulations whose exact inner-outer pairing was never in the training set, eliminating the need to enumerate all |Y_in| × |Y_out| combinations.
  • Adding a new modulation type to either layer requires only registering a new prototype in the disentangled semantic space, not retraining the network.
  • The logarithmic linearization provides a principled robustness route: since hardware and channel distortions also enter multiplicatively, they too are linearized by the same projection and can be handled as additive nuisance terms.
  • The reported robustness down to 4 dB SNR under combined impairments suggests the method is usable in non-cooperative settings such as spectrum monitoring and electronic intelligence, where parameters are unknown a priori.

Where Pith is reading between the lines

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

  • The homomorphism is defined for strictly positive real waveforms, but the paper applies it to complex baseband signals without specifying the branch of the complex logarithm or restricting to log-magnitude; if the complex extension is not well-defined, the claimed linearization is not mathematically established for the actual signal model, though the network might still learn a functional approxim
  • A direct test of the mechanism: replace the complex log with a fixed-branch log or log-magnitude-plus-phase and re-run TSDN; if zero-shot accuracy is unchanged, the logarithmic linearization is not the operative cause of the improvement.
  • The failure-mode analysis shows that a small known-class ambiguity (QPSK vs 8PSK) cascades into severe zero-shot misclassification when the outer layer is hard to disentangle; this implies that per-layer prototype margin and outer-layer separation should be optimized jointly, a coupling the paper does not exploit.
  • The high variance in ablations where the log mapping is removed (std ≈ 31%) suggests the mechanism may be fragile for certain held-out pairs; a per-pair error analysis beyond the three reported cases would clarify which combinations the linearization does not cover.

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 / 4 minor

Summary. The manuscript proposes a compositional zero-shot recognition framework, TSDN, for automatic composite modulation recognition. The signal model writes the received waveform as a product of an inner-layer and an outer-layer modulation, plus multiplicative hardware impairments and additive noise. The authors propose to apply an element-wise logarithm to map the received signal into a 'tangent space,' then use a spatial transformer network to learn input-dependent affine transformations that disentangle the two layers, and finally match the resulting layer-wise features against learned prototypes. Experiments on a self-generated dataset report 93.52% unknown-class accuracy for the BPSK-LFM holdout, along with ablations and robustness studies under AWGN, multipath, and hardware impairments.

Significance. If the central claims were established, the paper would make a useful contribution: a lightweight architecture with compositional zero-shot generalization to unseen inner-outer modulation pairs would be of clear interest to ACMR/ISAC. The paper also contains a well-structured problem formulation, a documented signal model, systematic ablations, and a per-holdout failure analysis, which are valuable. However, as detailed below, the theoretical linearization is not valid for the complex signal model actually used, Proposition 1 is a tautology, and the reported empirical results are internally inconsistent. The central claims are therefore not reliable as they stand.

major comments (4)
  1. [III-A, Eqs. (20)-(22)] The central linearization x = x_CM + x_Delta + n_tilde is obtained by applying an element-wise logarithm to the received signal s_rx. But s_rx, s_CM, Delta_s, and eta are complex-valued baseband samples (Eqs. (1)-(15)); the complex logarithm is multi-valued modulo 2*pi*i and undefined at zeros. The paper neither restricts to log-magnitude nor specifies a phase branch. The stated homomorphism log:(R_+,x)->(R,+) cannot be applied to the complex signal model, so the additive decomposition (22) is not mathematically established. Since the entire disentangled semantic space and the STN operate on this decomposition, the foundation of the method is unsupported.
  2. [III-B, Prop. 1 and Eq. (24)] Proposition 1 asserts that an input-dependent affine map perfectly isolates x_CM. The proof chooses A* = c_theta I and b* = -c_theta(x_Delta + n_tilde); substitution makes Eq. (24) true by construction. However, x_Delta and n_tilde are unobservable nuisance terms, and no estimation procedure is given. As stated, the proposition is therefore a restatement of the definition rather than a derivable guarantee; it does not support the claim that an STN can learn the ideal transform from data. A nontrivial identifiability or consistency result, or a constructive estimator, is needed.
  3. [V-G-2, Figs. 11-12; Abstract] The abstract's 'over 93% zero-shot recognition accuracy' is not representative. The 93.52% figure is reported for the single easiest holdout, BPSK-LFM (Table IV and Dataset 1 in Table V). The paper's own per-holdout analysis shows that when QPSK-LFM is held out, 88.8% of QPSK samples are misclassified as 8PSK in Branch 1, so the unknown composite accuracy for that cell is at most about 11%; for QPSK-MSK, 98.6% QPSK misclassification gives at most about 1.4%; for 32QAM-MSK, 66.9% MSK-to-NONE misclassification gives at most about 33%. These values are far below the headline and contradict the claim that TSDN generalizes reliably to unseen composite modulations. Section VI, item 3, itself acknowledges the non-uniform recognition performance, but the abstract's claim is unqualified.
  4. [Tables II and IV] The reported statistics are internally impossible or unreliable. In Table II, unknown-set entries such as 1.0000±1.0000 and 0.1507±0.6473 exceed the [0,1] bounds for accuracy/F1, which cannot occur as mean±std. In Table IV, unknown accuracy 0.8351±0.3135 has a negative 1-sigma lower bound, and the '95% confidence interval' is not defined; the full model's unknown accuracy 0.9352±0.0310 (std 3.10%) is hard to reconcile with the w/o Log row's 31.35% std despite only a 5.6% mean drop. These issues undermine the quantitative basis of the empirical claims.
minor comments (4)
  1. [Eq. (20)] The definition of eta involves division by s_CM*Delta_s; this ratio is undefined at samples where the product is zero (which can occur for QAM/PSK pulses). The paper should state the required support assumptions.
  2. [Eq. (28)] The dimensions of [x 1] and Theta_in are not specified. Proposition 1 uses a 2xN parameter matrix, while Eq. (28) writes [x 1]Theta with an N x (N+1) input; please define the exact shapes and how the log-magnitude and phase components are fed into the STN.
  3. [Fig. 9 and V-G-1] The text states that the full model 'successfully maintains high zero-shot accuracy' on Dataset 3 (128QAM-8FSK), but no numerical value is given in the figure or text. Please report the actual accuracy and reconcile it with the severe failures shown in Fig. 11.
  4. [Table IV] The '95% confidence interval' label is not justified for three random seeds. Please specify the interval construction or replace it with the standard deviation of the reported mean.

Circularity Check

1 steps flagged

The core affine-disentanglement 'sufficiency' proof is tautological—the affine parameters are constructed from the very nuisance components they are supposed to remove—but the zero-shot holdout evaluation itself is a genuine independent test, so circularity is partial.

specific steps
  1. self definitional [Section III-B, Proposition 1, Eqs. (23)–(24)]
    "Setting A∗ = cθ I and b∗ = −cθ(x∆ + ˜n) and substituting x = xCM + x∆ + ˜n yields: xA∗+b∗ = cθ(xCM+x∆+˜n)−cθ(x∆+˜n) = cθ xCM."

    The proof constructs the affine parameters A*, b* from xΔ+ñ, the exact nuisance components the transformation is supposed to eliminate. Those components are unobservable and not uniquely determined by x, so Eq. (24) is an algebraic identity: b* is chosen to cancel whatever is not xCM. The proposition therefore restates that an additive mixture can be unmixed if the component to be removed is already known; it does not establish that any input-dependent affine map computable from x alone isolates xCM. The claimed 'sufficiency' reduces to its own construction.

full rationale

The central theoretical step in Section III-B is circular: Proposition 1's 'Affine Disentanglement Sufficiency' is proven by defining the affine bias as the negative of the nuisance terms xΔ+ñ, so the equality in Eq. (24) holds by construction rather than by derivation. The paper concedes 'the oracle parameters depend on the unobservable components,' which makes the existence result vacuous as a justification for the learnable disentangling transform. This warrants a partial-circularity score. However, the empirical zero-shot evaluation is a genuine holdout: unseen composite pairs such as BPSK-LFM are excluded from training while their individual layer labels are seen, and the learned prototypes are fitted only on seen data; the reported 93% is not forced by the training labels in the sense of a fitted quantity renamed as a prediction. The abstract's 'over 93%' claim is nevertheless non-representative: Section V-G-2 reports catastrophic per-holdout failures (88.8% of QPSK misclassified as 8PSK for QPSK-LFM, 98.6% for QPSK-MSK, 66.9% of MSK misclassified as NONE for 32QAM-MSK), and the conclusion itself admits 'Our analysis reveals non-uniform recognition performance across different CM signal combinations.' That is a statistical/cherry-picking problem, not a circular derivation, so it is noted but not counted as a second circular step. The self-citation [49] used as 'OPT-6.7B' in the unified-semantic baseline is a reference error and is not load-bearing: the baseline's failure is demonstrated by the paper's own experiments rather than by the citation. The complex-logarithm issue (T(·)=log(·) applied to complex baseband signals) is a correctness risk, not circularity. Thus the only concrete circular step is the tautological Proposition 1.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

No new physical entities are introduced. The scientific load is carried by the multiplicative CM model, the multiplicative-hardware model, and the (invalid for complex signals) log homomorphism, plus the empirical learnability of the STN-based affine separation.

free parameters (3)
  • Loss weights λ1, λ2, λ3 = λ1=1, λ2=0.002, λ3=0.001
    Hand-chosen; Table IV ablations show L3 strongly affects unknown accuracy, so these weights are load-bearing.
  • Backbone depth (number of 1D conv layers) = unspecified; 3–4 layers saturate per Fig. 9
    Deep enough to handle high-order modulations; exact final value used for the headline 93% result is not stated.
  • Feature dimension, batch size, initial LR, epochs = d=64, batch=128, lr=1e-3, 200 epochs
    Standard training hyperparameters; not central to the claim but needed for replication.
axioms (4)
  • domain assumption CM signal is the element-wise product s_CM = s_in ⊙ s_out (Eq. 8)
    The whole framework relies on multiplicative coupling at the waveform level; this is stated as the CM model in Section II-A.
  • domain assumption Hardware imperfections collapse into a pointwise multiplicative distortion Δs(n) (Eq. 10)
    Assumed so that the log transform linearizes nuisance terms; follows from [31] but is a modeling choice.
  • domain assumption T(·)=log(·) is a valid linearization for the received complex signal s_rx
    Section III-A applies the real-positive log homomorphism to complex baseband signals without specifying branch cuts or zero handling; this is the central unsupported assumption.
  • domain assumption A network trained on seen CM pairs learns affine parameters that generalize to unseen pairs
    Proposition 1 only gives an existence result using oracle parameters; the learnability of Θ(x)=T_Φ(x) from data is an empirical assumption, contradicted by the failure cases in Fig. 11.

pith-pipeline@v1.3.0-alltime-deepseek · 21100 in / 10955 out tokens · 111407 ms · 2026-08-02T05:06:20.539370+00:00 · methodology

0 comments
read the original abstract

Automatic composite modulation recognition (ACMR) is critical for integrated sensing and communication (ISAC) systems, while conventional approaches face significant challenges due to the semantic coupling between inner-layer and outer-layer modulations in composite modulation (CM), degraded performance under joint hardware and channel imperfections, and limited capability to handle unknown modulation schemes. To this end, we design a disentangled semantic space and propose zero-shot learning framework. Within this framework, a logarithmic projection first linearizes the multiplicative coupling between modulation layers and a learnable geometric transformation is used for layer-wise semantic features. We instantiate the framework as the Tangent Space Disentanglement Network (TSDN). TSDN integrates logarithmic mapping, a spatial transformer network for learning the geometric transformation, and a multi-objective loss function that balances discrimination with cross-domain generalization. Comprehensive experiments demonstrate that TSDN achieves over 93\% zero-shot recognition accuracy, outperforms unified-semantic and multi-task baselines by significant margins, and maintains robust performance under combined channel fading and hardware imperfections down to 4 dB SNR.

Figures

Figures reproduced from arXiv: 2607.13463 by Baoguo Li, Xiang Wang, Yurui Zhao, Zhitao Huang.

Figure 1
Figure 1. Figure 1: Schematic of the ISAC system showing the role of composite modulation (CM) in enabling sensing and communication, with applications in vehicle [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Structure of an ISAC emitter with CM modulator contains two-layer [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Time-domain waveforms (top row) and time-frequency spectrograms (bottom row) of CM signals illustrate how the inner-layer symbols ride on the outer-layer chirp structure. (a),(d) Pure inner-layer modulation (BPSK). (b),(d) Pure outer-layer modulation (LFM). (c),(f) Composite modulation (BPSK-LFM). B. Imperfection Factors In practice, the received signal deviates from the ideal CM waveform due to transmitte… view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of zero-shot learning frameworks. (a) Traditional unified [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Architecture of the proposed TSDN for zero-shot automatic composite modulation recognition based on tangent space disentanglement. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of semantic features for known signals. The plot includes 25 composite modulation signals and 10 single modulation signals. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Scatter plot comparing model performance on known and unknown [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of different neural network architectures. (a) Performance [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Model performance across high-order modulations with varying [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Distribution of known-class and unknown-class accuracy across all [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Per-branch accuracy heatmap at backbone depth = 3. Each row corresponds to one CM type held out as unseen. Columns show the accuracy [PITH_FULL_IMAGE:figures/full_fig_p014_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Confusion matrices for the three most challenging held-out cases. (a)(d) Branch 1 when QPSK-LFM is unseen: QPSK-8PSK confusion. (b)(e) [PITH_FULL_IMAGE:figures/full_fig_p014_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Known-class and unknown-class accuracy versus SNR under [PITH_FULL_IMAGE:figures/full_fig_p015_13.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

61 extracted references · 2 linked inside Pith

  1. [1]

    Radio resource management in joint radar and communication: A comprehen- sive survey,

    N. C. Luong, X. Lu, D. T. Hoang, D. Niyato, and D. I. Kim, “Radio resource management in joint radar and communication: A comprehen- sive survey,”IEEE Communications Surveys & Tutorials, vol. 23, no. 2, pp. 780–814, 2021

    2021

  2. [2]

    A survey on fundamental limits of integrated sensing and communication,

    A. Liu, Z. Huang, M. Li, Y . Wan, W. Li, T. X. Han, C. Liu, R. Du, D. K. P. Tan, J. Luet al., “A survey on fundamental limits of integrated sensing and communication,”IEEE Communications Surveys & Tutorials, vol. 24, no. 2, pp. 994–1034, 2022

    2022

  3. [3]

    Radar and communi- cation coexistence: An overview: A review of recent methods,

    L. Zheng, M. Lops, Y . C. Eldar, and X. Wang, “Radar and communi- cation coexistence: An overview: A review of recent methods,”IEEE Signal Processing Magazine, vol. 36, no. 5, pp. 85–99, 2019

    2019

  4. [4]

    Integrated sensing and communication waveform design: A survey,

    W. Zhou, R. Zhang, G. Chen, and W. Wu, “Integrated sensing and communication waveform design: A survey,”IEEE Open Journal of the Communications Society, vol. 3, pp. 1930–1949, 2022

    1930

  5. [5]

    A unified distributed algorithm for hybrid near-far field activity detection in cell- free massive mimo,

    J. Lei, Y . Li, Z. Wang, Q. Lin, Y .-F. Liu, and Y .-C. Wu, “A unified distributed algorithm for hybrid near-far field activity detection in cell- free massive mimo,”arXiv preprint arXiv:2509.15162, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    Novel radar waveform optimization for a cooperative radar-communications system,

    A. R. Chiriyath, S. Ragi, H. D. Mittelmann, and D. W. Bliss, “Novel radar waveform optimization for a cooperative radar-communications system,”IEEE Transactions on Aerospace and Electronic Systems, vol. 55, no. 3, pp. 1160–1173, 2019

    2019

  7. [7]

    An overview of signal processing techniques for joint communication and radar sensing,

    J. A. Zhang, F. Liu, C. Masouros, R. W. Heath, Z. Feng, L. Zheng, and A. Petropulu, “An overview of signal processing techniques for joint communication and radar sensing,”IEEE Journal of Selected Topics in Signal Processing, vol. 15, no. 6, pp. 1295–1315, 2021

    2021

  8. [8]

    Design of frequency index modulated waveforms for integrated sar and communication on high- altitude platforms (haps),

    B. Huang, S. Ahmed, and M.-S. Alouini, “Design of frequency index modulated waveforms for integrated sar and communication on high- altitude platforms (haps),”IEEE Transactions on Communications, 2025

    2025

  9. [9]

    Automatic waveform recognition of overlapping lpi radar signals based on multi-instance multi-label learning,

    Z. Pan, S. Wang, M. Zhu, and Y . Li, “Automatic waveform recognition of overlapping lpi radar signals based on multi-instance multi-label learning,”IEEE Signal Processing Letters, vol. 27, pp. 1275–1279, 2020

    2020

  10. [10]

    Intra-pulse modulation recognition of dual- component radar signals based on deep convolutional neural network,

    W. Si, C. Wan, and Z. Deng, “Intra-pulse modulation recognition of dual- component radar signals based on deep convolutional neural network,” IEEE Communications Letters, vol. 25, no. 10, pp. 3305–3309, 2021

    2021

  11. [11]

    Likelihood methods for mpsk modulation classification,

    C.-Y . Huan and A. Polydoros, “Likelihood methods for mpsk modulation classification,”IEEE Transactions on Communications, vol. 43, no. 2/3/4, pp. 1493–1504, 1995

    1995

  12. [12]

    A survey on machine- learning techniques in cognitive radios,

    M. Bkassiny, Y . Li, and S. K. Jayaweera, “A survey on machine- learning techniques in cognitive radios,”IEEE Communications Surveys & Tutorials, vol. 15, no. 3, pp. 1136–1159, 2012

    2012

  13. [13]

    Convolutional radio mod- ulation recognition networks,

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

    2016

  14. [14]

    Deep architectures for modulation recog- nition,

    N. E. West and T. O’shea, “Deep architectures for modulation recog- nition,” in2017 IEEE international symposium on dynamic spectrum access networks (DySPAN). IEEE, 2017, pp. 1–6

    2017

  15. [15]

    Cnn-based automatic modulation classification for beyond 5g communications,

    A. P. Hermawan, R. R. Ginanjar, D.-S. Kim, and J.-M. Lee, “Cnn-based automatic modulation classification for beyond 5g communications,” IEEE Communications Letters, vol. 24, no. 5, pp. 1038–1041, 2020

    2020

  16. [16]

    Modulation recognition with pre-denoising convolu- tional neural network,

    Y . Liu and Y . Liu, “Modulation recognition with pre-denoising convolu- tional neural network,”Electronics Letters, vol. 56, no. 5, pp. 255–257, 2020

    2020

  17. [17]

    Understanding complex-valued transformer for modulation recognition,

    J. Lei, Y . Li, L.-Y . Yung, Y . Leng, Q. Lin, and Y .-C. Wu, “Understanding complex-valued transformer for modulation recognition,”IEEE Wireless Communications Letters, vol. 13, no. 12, pp. 3523–3527, 2024

    2024

  18. [18]

    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 Transactions on Cognitive Communica- tions and Networking, vol. 4, no. 3, pp. 433–445, 2018

    2018

  19. [19]

    Real-time radio technology and modulation classification via an lstm auto-encoder,

    Z. Ke and H. Vikalo, “Real-time radio technology and modulation classification via an lstm auto-encoder,”IEEE Transactions on Wireless Communications, vol. 21, no. 1, pp. 370–382, 2021

    2021

  20. [20]

    Automatic modulation classification using cnn-lstm based dual-stream structure,

    Z. Zhang, H. Luo, C. Wang, C. Gan, and Y . Xiang, “Automatic modulation classification using cnn-lstm based dual-stream structure,” IEEE Transactions on Vehicular Technology, vol. 69, no. 11, pp. 13 521– 13 531, 2020

    2020

  21. [21]

    Multidimensional cnn-lstm network for automatic modulation classification,

    N. Wang, Y . Liu, L. Ma, Y . Yang, and H. Wang, “Multidimensional cnn-lstm network for automatic modulation classification,”Electronics, vol. 10, no. 14, p. 1649, 2021

    2021

  22. [22]

    Convolutional, long short-term memory, fully connected deep neural networks,

    T. N. Sainath, O. Vinyals, A. Senior, and H. Sak, “Convolutional, long short-term memory, fully connected deep neural networks,” in 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP). Ieee, 2015, pp. 4580–4584

    2015

  23. [23]

    Modulation recognition of composite modulation signal based on two-fold digital receiver and goodness of fit test,

    W. Lijun, H. Yu, Z. Pan, G. Hongfang, and S. Lei, “Modulation recognition of composite modulation signal based on two-fold digital receiver and goodness of fit test,” in2021 IEEE International Conference on Electronic Technology, Communication and Information (ICETCI). IEEE, 2021, pp. 433–437

    2021

  24. [24]

    Automatic composite-modulation classification using cyclic-paw-print features for cognitive aerospace communications,

    X. Yan, X. Zhong, H.-C. Wu, P. Yang, Q. Wang, and Y . Chen, “Automatic composite-modulation classification using cyclic-paw-print features for cognitive aerospace communications,”IEEE Transactions on Communications, vol. 72, no. 9, pp. 5486–5502, 2024

    2024

  25. [25]

    Efficient auto- matic composite-modulation classifier using cyclic-paw-print features,

    X. Yan, Y . Chen, X. Zhong, H.-C. Wu, and Q. Wang, “Efficient auto- matic composite-modulation classifier using cyclic-paw-print features,” IEEE Communications Letters, vol. 28, no. 3, pp. 652–656, 2024

    2024

  26. [26]

    Auto- matic composite-modulation classification using ultra lightweight deep- learning network based on cyclic-paw-print,

    X. Yan, P. Yang, X. Zhong, Q. Wang, H.-C. Wu, and L. He, “Auto- matic composite-modulation classification using ultra lightweight deep- learning network based on cyclic-paw-print,”IEEE Transactions on Cognitive Communications and Networking, vol. 10, no. 3, pp. 866– 879, 2024

    2024

  27. [27]

    Modulation recognition of composite signal based on resnet and frequency domain graph,

    L. Wang, Y . Han, H. Ge, Y . Zhao, and L. Shen, “Modulation recognition of composite signal based on resnet and frequency domain graph,” in Journal of Physics: Conference Series, vol. 2025, no. 1. IOP Publishing, 2021, p. 012029

    2025

  28. [28]

    Blind recognition algorithm of multi-carrier composite modulation signal based on multi- dimensional time-frequency superimposed spectrum,

    S. Wang, H. Li, X. Zhang, H. Jiang, and L. Shen, “Blind recognition algorithm of multi-carrier composite modulation signal based on multi- dimensional time-frequency superimposed spectrum,”Sensors, vol. 25, no. 13, p. 4007, 2025

    2025

  29. [29]

    Composite radar modulation identification by transfer learning,

    F. Li, Z. Yang, B. Huang, and Y . Chen, “Composite radar modulation identification by transfer learning,” in2021 International Conference on Microwave and Millimeter Wave Technology (ICMMT). IEEE, 2021, pp. 1–3

    2021

  30. [30]

    Radio frequency and modulation systems-part 1 earth stations and space craft,

    C. C. for Space Data Systems (CCSDSet al., “Radio frequency and modulation systems-part 1 earth stations and space craft,”Recommended standard CCSDS 401.0-B, 2009. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 17

    2009

  31. [31]

    Concentrate on hardware imperfec- tion via aligning reconstructed states,

    Y . Zhao, X. Wang, and Z. Huang, “Concentrate on hardware imperfec- tion via aligning reconstructed states,”IEEE Communications Letters, vol. 26, no. 12, pp. 2934–2938, 2022

    2022

  32. [32]

    Prediction of performance loss due to phase noise in digital satellite communication system,

    Y .-w. Kim and D.-c. Park, “Prediction of performance loss due to phase noise in digital satellite communication system,” inIST Mobile & Wireless Telecommunications conf.Citeseer, 2002, pp. 575–578

    2002

  33. [33]

    Specific emitter identification using nonlinear device estimation,

    M.-W. Liu and J. F. Doherty, “Specific emitter identification using nonlinear device estimation,” in2008 IEEE Sarnoff Symposium. IEEE, 2008, pp. 1–5

    2008

  34. [34]

    Absil, R

    P.-A. Absil, R. Mahony, and R. Sepulchre,Optimization Algorithms on Matrix Manifolds. Princeton University Press, 2008

    2008

  35. [35]

    Spatial transformer networks,

    M. Jaderberg, K. Simonyan, A. Zisserman, and k. kavukcuoglu, “Spatial transformer networks,” inAdvances in Neural Information Processing Systems, C. Cortes, N. Lawrence, D. Lee, M. Sugiyama, and R. Garnett, Eds., vol. 28. Curran Associates, Inc., 2015

    2015

  36. [36]

    Cross-entropy loss functions: Theoretical analysis and applications,

    A. Mao, M. Mohri, and Y . Zhong, “Cross-entropy loss functions: Theoretical analysis and applications,” inInternational conference on Machine learning. pmlr, 2023, pp. 23 803–23 828

    2023

  37. [37]

    A discriminative feature learning approach for deep face recognition,

    Y . Wen, K. Zhang, Z. Li, and Y . Qiao, “A discriminative feature learning approach for deep face recognition,” inECCV, 2016

    2016

  38. [38]

    Domain separation networks,

    K. Bousmalis, G. Trigeorgis, N. Silberman, D. Krishnan, and D. Erhan, “Domain separation networks,” inAdvances in Neural Information Processing Systems, D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett, Eds., vol. 29. Curran Associates, Inc., 2016

    2016

  39. [39]

    Cognitive radio for satellite tt & c system: A general dataset using software-defined radio,

    Y . Zhang, B. Zang, H. Ji, L. Li, S. Li, and L. Chen, “Cognitive radio for satellite tt & c system: A general dataset using software-defined radio,” Scientific Data, 2026

    2026

  40. [40]

    Hisarmod: A new challenging modulated signals dataset,

    K. Tekbıyık, C. Kec ¸eci, A. Ekti, A. G ¨orc ¸in, and G. Kurt, “Hisarmod: A new challenging modulated signals dataset,”IEEE Dataport, 2019

    2019

  41. [41]

    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

  42. [42]

    Multi- stage learning for radar pulse activity segmentation,

    Z. Huang, A. Pemasiri, S. Denman, C. Fookes, and T. Martin, “Multi- stage learning for radar pulse activity segmentation,” inICASSP 2024- 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 7340–7344

    2024

  43. [43]

    Robust and fast automatic modulation classification with cnn under multipath fading channels,

    K. Tekbıyık, A. R. Ekti, A. G ¨orc ¸in, G. K. Kurt, and C. Kec ¸eci, “Robust and fast automatic modulation classification with cnn under multipath fading channels,” in2020 IEEE 91st Vehicular Technology Conference (VTC2020-Spring). IEEE, 2020, pp. 1–6

    2020

  44. [44]

    Mcnet: An efficient cnn architecture for robust automatic modulation classification,

    T. Huynh-The, C.-H. Hua, Q.-V . Pham, and D.-S. Kim, “Mcnet: An efficient cnn architecture for robust automatic modulation classification,” IEEE Communications Letters, vol. 24, no. 4, pp. 811–815, 2020

    2020

  45. [45]

    Deep neural network architectures for modulation classification,

    X. Liu, D. Yang, and A. El Gamal, “Deep neural network architectures for modulation classification,” in2017 51st Asilomar Conference on Signals, Systems, and Computers. IEEE, 2017, pp. 915–919

    2017

  46. [46]

    Automatic modulation classification using recurrent neural networks,

    D. Hong, Z. Zhang, and X. Xu, “Automatic modulation classification using recurrent neural networks,” in2017 3rd IEEE international conference on computer and communications (ICCC). IEEE, 2017, pp. 695–700

    2017

  47. [47]

    A hierarchical classification head based convolutional gated deep neural network for automatic modulation classification,

    S. Chang, R. Zhang, K. Ji, S. Huang, and Z. Feng, “A hierarchical classification head based convolutional gated deep neural network for automatic modulation classification,”IEEE Transactions on Wireless Communications, vol. 21, no. 10, pp. 8713–8728, 2022

    2022

  48. [48]

    Visualizing data using t-sne

    L. Van der Maaten and G. Hinton, “Visualizing data using t-sne.”Journal of machine learning research, vol. 9, no. 11, 2008

    2008

  49. [49]

    Zero-shot automatic modulation recognition using a large vision-language model,

    Y . Zhao, X. Wang, S. Cao, and Z. Huang, “Zero-shot automatic modulation recognition using a large vision-language model,”IEEE Transactions on Communications, vol. 73, pp. 15 765–15 782,

  50. [50]

    Sr2cnn: Zero- shot learning for signal recognition,

    Y . Dong, X. Jiang, H. Zhou, Y . Lin, and Q. Shi, “Sr2cnn: Zero- shot learning for signal recognition,”IEEE Transactions on Signal Processing, vol. 69, pp. 2316–2329, 2021

    2021

  51. [51]

    Deep multi-task representation learning: A tensor factorisation approach,

    Y . Yang and T. M. Hospedales, “Deep multi-task representation learning: A tensor factorisation approach,” inICLR, 2017

    2017

  52. [52]

    Multi-task learning using uncer- tainty to weigh losses for scene geometry and semantics,

    A. Kendall, Y . Gal, and R. Cipolla, “Multi-task learning using uncer- tainty to weigh losses for scene geometry and semantics,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 7482–7491

    2018

  53. [53]

    Physics-informed generalizable wireless channel modeling with segmentation and deep learning: Fundamen- tals, methodologies, and challenges,

    E. Zhu, H. Sun, and M. Ji, “Physics-informed generalizable wireless channel modeling with segmentation and deep learning: Fundamen- tals, methodologies, and challenges,”IEEE Wireless Communications, vol. 31, no. 6, pp. 170–177, 2024

    2024

  54. [54]

    Physics-integrated inference for signal recovery in non-gaussian regimes,

    M. A. Mousa, L. Bauer, Z. Yang, U. Singh, A. Deka, and Z. Ja- cob, “Physics-integrated inference for signal recovery in non-gaussian regimes,”arXiv preprint arXiv:2601.18074, 2026

    arXiv 2026

  55. [55]

    Large lan- guage models are zero-shot reasoners,

    T. Kojima, S. S. Gu, M. Reid, Y . Matsuo, and Y . Iwasawa, “Large lan- guage models are zero-shot reasoners,”Advances in neural information processing systems, vol. 35, pp. 22 199–22 213, 2022

    2022

  56. [56]

    Region graph embedding network for zero-shot learning,

    G.-S. Xie, L. Liu, F. Zhu, F. Zhao, Z. Zhang, Y . Yao, J. Qin, and L. Shao, “Region graph embedding network for zero-shot learning,” inEuropean conference on computer vision. Springer, 2020, pp. 562–580

    2020

  57. [57]

    Data imbal- ance in classification: Experimental evaluation,

    F. Thabtah, S. Hammoud, F. Kamalov, and A. Gonsalves, “Data imbal- ance in classification: Experimental evaluation,”Information Sciences, vol. 513, pp. 429–441, 2020

    2020

  58. [58]

    Dynamic model pruning with feedback,

    T. Lin, S. U. Stich, L. Barba, D. Dmitriev, and M. Jaggi, “Dynamic model pruning with feedback,”arXiv preprint arXiv:2006.07253, 2020

    Pith/arXiv arXiv 2006

  59. [59]

    Pruning vs quantization: Which is better?

    A. Kuzmin, M. Nagel, M. Van Baalen, A. Behboodi, and T. Blankevoort, “Pruning vs quantization: Which is better?”Advances in neural infor- mation processing systems, vol. 36, pp. 62 414–62 427, 2023

    2023

  60. [60]

    Analysis of programs for parallel processing,

    A. J. Bernstein, “Analysis of programs for parallel processing,”IEEE transactions on electronic computers, no. 5, pp. 757–763, 1966

    1966

  61. [2025]

    Available: https://api.semanticscholar.org/CorpusID: 281447064

    [Online]. Available: https://api.semanticscholar.org/CorpusID: 281447064