Pith. sign in

REVIEW 4 major objections 6 minor 58 references

Task-Oriented Low-Label Semantic Communication With Self-Supervised Learning

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that self-supervised pre-training of a semantic encoder on unlabeled data, followed by end-to-end joint source-channel coding fine-tuning with few labels, makes task-oriented semantic communication accurate over noisy…

desk verdict SSL pre-training helps low-label semantic communication, but the 'irrelevant unlabeled data' claim is overbroad and unverified; worth a serious referee after revisions. read the letter →

arxiv 2505.19940 v1 pith:JZN5HT7F submitted 2025-05-26 cs.LG eess.SP

Run Gu , Wei Xu , Zhaohui Yang , Dusit Niyato , Aylin Yener This is my paper
classification cs.LGeess.SP
keywords semanticcommunicationtask-orientedself-supervisedlearninginformationbottleneckjointsource-channelcodinglow-labelcontrastive
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 proposes a semantic communication system for image classification when labeled training data is scarce. The authors claim that pre-training a semantic encoder on unlabeled images with self-supervised contrastive and reconstruction tasks, then fine-tuning it end-to-end with a joint source-channel coding (JSCC) network using a small labeled set, yields more accurate task inference over multipath fading channels than digital coding, training from scratch, or transferring a pre-trained model. The central claim is that the unlabeled data supplies task-relevant semantic knowledge through an information-bottleneck-style objective, and that the improvement persists at low SNR and even when the unlabeled data comes from a different distribution. A reader would care because labels are the expensive bottleneck in real edge networks, and this work shifts most of the learning burden to cheap unlabeled data.

What carries the argument

The load-bearing mechanism is the reformulation of the information bottleneck objective into two pretext tasks on unlabeled data: an InfoNCE classification task that maximizes a lower bound on $I(X;R)$ between extracted semantic features and augmented views, and a reconstruction task that approximates minimizing $H(X|R)$ by penalizing $\|x - x_r\|_2^2$ under a Gaussian conditional model. Their weighted sum $L_{\mathrm{pre}} = -(\lambda L_r + L_c)$ trains the semantic encoder to keep task-relevant shared content while discarding augmentation-induced irrelevant variation. The auxiliary projection used for contrastive similarity is discarded after pre-training.

What would settle it

Run the same two-stage pipeline but replace augmentations with content-destroying ones, such as crops containing no object or heavy masking, and use unlabeled data whose classes do not overlap the downstream labels; if the gains over training from scratch vanish, the multi-view assumption is load-bearing, and if the gains persist, the assumption is unnecessary.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that a two-stage learning framework, called SLSCom, improves end-to-end classification accuracy in task-oriented semantic communication under label scarcity. In the first stage, the semantic encoder is trained locally on unlabeled images by maximizing an InfoNCE classification objective, a lower bound on mutual information between the extracted feature and a randomly augmented view, and by a reconstruction objective that approximates minimizing conditional entropy, together forming a tractable surrogate for the information bottleneck problem. In the second stage, the pre-trained encoder is integrated with JSCC encoding and decoding and a task decoder over OFDM multipath channels, either frozen or fine-tuned with few labeled samples. The reported simulations show consistent gains over digital coding, training from scratch, and transfer learning across labeled set sizes, SNR values, partial label missing, and distribution-shifted unlabeled data, with the largest gains in low-label, low-SNR regimes.

Load-bearing premise

The framework assumes that random augmentations of an unlabeled image keep enough task-relevant content, formalized as a Markov chain plus an epsilon lower bound on shared information, so that contrastive features learned from unlabeled data are actually useful for the downstream label.

Editorial extensions

If this is right

  • A device can build a useful task-oriented semantic encoder from data it already collects without labels, before any annotated samples arrive.
  • With as few as 2,000 labeled images, the proposed framework's classification accuracy exceeds training from scratch by a wide margin at 0 dB SNR.
  • Pre-training on unlabeled data from another domain, such as SVHN for CIFAR-10, still helps, indicating that the learned semantic extraction is not tightly bound to the exact downstream distribution.
  • The JSCC network contributes roughly a 7 dB SNR gain at 80 percent accuracy over a hybrid semantic and digital baseline, meaning that the jointly trained transmission module is a major source of robustness.
  • The reconstruction pretext task matters most for lighter networks with smaller feature dimensions, where discarding task-irrelevant information is essential.

Reading between the lines

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

  • If the multi-view assumption is the real source of gain, then choosing augmentations that deliberately preserve semantic content while varying style, color, and geometry should matter more than the amount of unlabeled data; the paper's ablations point this way but do not isolate it.
  • The same two-stage recipe could be tested on detection or segmentation tasks, since the semantic encoder is task-agnostic and only the decoder changes; the paper states such generality but only evaluates classification.
  • A straightforward extension is to make the unlabeled set class-balanced or actively selected, which should further improve low-label performance beyond the uniform-split gains reported.
  • Training the semantic encoder locally and discarding the auxiliary projection suggests a practical deployment pattern in which edge devices update their semantic encoders incrementally without retraining the channel codec; the paper mentions multi-device extension as future work.
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

4 major / 6 minor

Summary. The paper proposes SLSCom, a two-stage learning framework for task-oriented semantic communication over wireless channels under limited labeled data. In the first stage, a semantic encoder is pre-trained locally on unlabeled data using a contrastive InfoNCE loss and a reconstruction loss, motivated by an information bottleneck (IB) reformulation. In the second stage, the encoder is fine-tuned (or frozen) together with a JSCC network and a semantic decoder using a small labeled set. The paper reports simulations on CIFAR10, SVHN, and Flowers showing that SLSCom outperforms training from scratch, ImageNet transfer, and conventional digital coding, especially at low SNR and with few labels.

Significance. The paper addresses a practically important problem—low-label task-oriented semantic communication—and provides a reasonably extensive empirical study: two network structures, multiple datasets, ablations for the reconstruction pretext task, auxiliary projection, color transformations, and robustness checks under distribution shift and partial label missing. The proposed method is clearly described and the simulation results are consistent with the expectation that self-supervised pre-training helps in low-label regimes. The main weaknesses are that the theoretical IB justification is asserted rather than proven, the approximations used to turn the IB objective into pretext tasks are heuristic, and the headline claim about robustness to task-irrelevant unlabeled data is not supported by the experiments (only a domain-shift experiment, SVHN-to-CIFAR10, is provided). These issues are load-bearing for the paper's central claims and need to be addressed before publication.

major comments (4)
  1. [Abstract and Section III-B (Eq. 10)] The abstract claims that SLSCom works 'even when the unlabeled samples are irrelevant to the downstream tasks,' but this claim is not supported by the experiments. The multi-view learning assumption in Section III-B requires I(Sa;T) ≥ I(R;T) ≥ I(Sa;R;T) ≥ ε; for truly task-irrelevant unlabeled data, I(Sa;T) would be near zero, violating this assumption. The only experiment that confronts this issue is SLSCom w DS (Section IV-D and Fig. 5), where unlabeled samples come from SVHN (digit classification) and labeled samples from CIFAR10 (object classification). SVHN and CIFAR10 are both natural 32×32 images with overlapping low-level statistics, so they are distribution-shifted but not task-irrelevant. The authors should either weaken the abstract and contribution statements to refer to 'distribution-shifted' unlabeled data, or add experiments with clearly task-irrelevant unlabeled data (e.g., random noise or images from unrelated label sets) to substantiate the stronger claim.
  2. [Section III-B] The assertion that the supervised IB problem (P1) and the self-supervised problem (P2) are 'equivalent' is not proven. (P1) optimizes I(X^;T) − λH(X^|T) using the task label T, whereas (P2) optimizes I(X;R) − λH(X|R) using the self-supervised view R. The text merely states that they share a 'common goal' and that the multi-view assumption ensures the shared information between Sa and R contains task-relevant components. No bound, equality, or formal connection between the two objectives is provided. This is a core theoretical claim, and it should be either derived under explicit assumptions or presented as a heuristic surrogate rather than an equivalence.
  3. [Section III-C (Eqs. 11-12)] The derivation of the pretext-task losses from the IB objective is heuristic. InfoNCE (Eq. 11) is a lower bound on mutual information I(X;R), and maximizing a lower bound is not in general equivalent to maximizing the true mutual information. The Gaussian approximation q(x|r) = CN(x|fse(r;Wse), I) used to turn H(X|R) into the L2 consistency loss in Eq. (12) is not justified for semantic features. The resulting objectives are plausible self-supervised losses, but the paper should explicitly acknowledge that they are only loosely connected to (P2) and not claim to 'solve' the IB problem exactly.
  4. [Section IV-A and all result figures/tables] The paper states that each approach is trained 10 times and the average is reported, but no standard deviations, confidence intervals, or error bars are shown anywhere (e.g., Fig. 5, Fig. 9, Table II). Given the repeated use of 'significantly outperforms,' the lack of variance information makes it impossible to assess whether the observed gaps, especially in the small-gap regimes at large labeled-set sizes, are statistically reliable. Please report means with error bars or confidence intervals.
minor comments (6)
  1. [Section IV-A] The sentence 'For Flower recognition dataset, both M1 and M are set to 2,814' is ambiguous because the experimental protocol uses M1 unlabeled samples for pre-training and M2 labeled samples for fine-tuning; please clarify how M, M1, and M2 relate for this dataset.
  2. [References] Reference [28] lists the arXiv ID 2407.03566 but the URL points to arXiv:1810.00740, which is a different paper; please correct the link.
  3. [Eq. (9)] The use of P and N to denote the sets of positive and negative pairs conflicts with the notation P1/P2 used for the optimization problems; consider using different symbols such as S+ and S−.
  4. [Fig. 11] The label 'Unfi.' in the figure legend is nonstandard; consider using 'Uniform' to denote the balanced per-class sampling.
  5. [Algorithm 2] In the parameter update for Wsd, the learning rate is written as 'µη' while all other updates use 'η'; this is mathematically correct (since the gradient of µLce is multiplied by η) but unconventional and may confuse readers—please factor the learning rate out or add a note.
  6. [Section III-C] The phrase 'The task aims to prompt the self-supervised data to reconstruct the semantic information' is awkward; consider rephrasing to 'The task aims to reconstruct the semantic information from the self-supervised data.'

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the IB reformulation, pretext-task surrogates, and multi-view assumption are explicit premises rather than disguised inputs, and all central claims are evaluated against independent baselines.

full rationale

The paper's derivation chain is self-contained in the sense required by the circularity analysis. The IB problem (P1) is reformulated as (P2) using self-supervision, and the pretext losses Lc (InfoNCE) and Lr (Gaussian reconstruction) are explicitly introduced as tractable approximations of I(X;R) and -H(X|R). This is a constructive approximation chain, not a case of fitting a parameter and then renaming the fit as a prediction. The load-bearing multi-view assumption is stated explicitly in Section III-B as I(Sa;T) >= I(R;T) >= I(Sa;R;T) >= epsilon and is cited to external multi-view learning literature [51]; the paper does not claim to derive this assumption from its own framework. The equality I(X*;T)=I(Sa;R;T) is asserted as a consequence of the multi-view assumption plus the IB optimization, and while it is not rigorously proven, it is a conditional claim rather than a definitional identity. Hyperparameters lambda and mu are tuned on validation sets, which is standard practice and does not constitute circular prediction. The only self-citation, [1], is a prior workshop presentation and is not used as load-bearing evidence. The advertised robustness to 'irrelevant' unlabeled data is supported only by a mild SVHN-to-CIFAR10 distribution shift, which is a claim-width or correctness concern, not a circularity concern. No prediction in the paper reduces by construction to a fitted value or to a self-citation chain, so the appropriate circularity score is 0.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the assumed multi-view relation between samples and augmentations and on an unproved equivalence between the supervised IB objective and the self-supervised objective. The loss functions used in practice are standard SSL losses repackaged as IB approximations.

free parameters (2)
  • IB weight lambda (λ) = 0.15 for CR_s=2/3, 0.1 for CR_s=1/6
    Selected by validation performance on the pretext classification task (Section IV-A-3), not derived.
  • Task loss weight mu (µ) = 1
    Chosen based on validation performance of downstream classification (Section IV-A-3).
assumptions (5)
  • domain assumption Multi-view learning assumption: unlabeled sample sa and augmented view r both contain sufficient task-relevant information, with Markov chain T ↔ Sa → R and I(Sa;T) ≥ I(R;T) ≥ I(Sa;R;T) ≥ ϵ.
    Invoked in Section III-B to justify that optimizing (P2) extracts task-relevant features. If augmentations lose task content, the pretext tasks are irrelevant to T.
  • ad hoc to paper Equivalence between the supervised IB problem (P1) and the self-supervised problem (P2).
    Asserted in Section III-B ('the equivalence between (P1) and (P2) lies in their common goal') without proof; this is a narrative bridge rather than a derived result.
  • ad hoc to paper Gaussian form q(x|r) = CN(x|fse(r;Wse), I) to approximate the conditional entropy H(X|R).
    Chosen in Section III-C-2 to convert the log-likelihood into an MSE loss; no justification beyond the Gaussian approximation used in image reconstruction literature.
  • domain assumption Data augmentation fdc preserves task-relevant content while changing style or appearance.
    Stated in Section III-B when designing transforms (random flip, ColorJitter, grayscale, Gaussian blur); the whole pre-training relies on this.
  • domain assumption Unlabeled samples are abundant relative to labeled ones (M1 is large, M2 small).
    The low-label scenario assumes M2 << M and M1 unlabeled samples are available (Section II-B, Section III-A).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Task-Oriented Low-Label Semantic Communication With Self-Supervised Learning." pith.science (2026). https://pith.science/paper/JZN5HT7F

@misc{pith2026250519940,
  author       = {Pith},
  title        = {Pith review of: Task-Oriented Low-Label Semantic Communication With Self-Supervised Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JZN5HT7F}},
  note         = {Machine review of arXiv:2505.19940}
}
read the original abstract

Task-oriented semantic communication enhances transmission efficiency by conveying semantic information rather than exact messages. Deep learning (DL)-based semantic communication can effectively cultivate the essential semantic knowledge for semantic extraction, transmission, and interpretation by leveraging massive labeled samples for downstream task training. In this paper, we propose a self-supervised learning-based semantic communication framework (SLSCom) to enhance task inference performance, particularly in scenarios with limited access to labeled samples. Specifically, we develop a task-relevant semantic encoder using unlabeled samples, which can be collected by devices in real-world edge networks. To facilitate task-relevant semantic extraction, we introduce self-supervision for learning contrastive features and formulate the information bottleneck (IB) problem to balance the tradeoff between the informativeness of the extracted features and task inference performance. Given the computational challenges of the IB problem, we devise a practical and effective solution by employing self-supervised classification and reconstruction pretext tasks. We further propose efficient joint training methods to enhance end-to-end inference accuracy over wireless channels, even with few labeled samples. We evaluate the proposed framework on image classification tasks over multipath wireless channels. Extensive simulation results demonstrate that SLSCom significantly outperforms conventional digital coding methods and existing DL-based approaches across varying labeled data set sizes and SNR conditions, even when the unlabeled samples are irrelevant to the downstream tasks.

Figures

Figures reproduced from arXiv: 2505.19940 by the authors.

Figure 1
Figure 1. Diagram of the proposed semantic communication. the necessity of the auxiliary projection layer in these challenging scenarios. The rest of this paper is organized as follows. Section II introduces the system model and presents design objectives of task-oriented semantic communication. Section III details the two￾stage learning framework and network structure of the proposed SLSCom. Simulation results are presented … view at source ↗
Figure 2
Figure 2. The framework and network overview of SLSCom. The parameters of fully-connected [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. A Venn diagram for (P2) with self-supervision. [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Pre-training for semantic encoder with pretext tasks. [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]
Figure 5
Figure 5. Figure 5: Comparison of test accuracy on CIFAR10 versus the number of labeled samples. [PITH_FULL_IMAGE:figures/full_fig_p022_5.png]
Figure 6
Figure 6. Figure 6: Training loss on CIFAR10 for RSCom, TSCom, and SLSCom with varying numbers of [PITH_FULL_IMAGE:figures/full_fig_p024_6.png]
Figure 7
Figure 7. Figure 7: Test accuracy comparison on CIFAR10 in partial label missing scenarios. [PITH_FULL_IMAGE:figures/full_fig_p025_7.png]
Figure 8
Figure 8. Figure 8: Heat map of test accuracy on SVHN versus the number of labeled samples. [PITH_FULL_IMAGE:figures/full_fig_p026_8.png]
Figure 9
Figure 9. Figure 9: Test accuracy comparison on CIFAR10 versus SNRs. [PITH_FULL_IMAGE:figures/full_fig_p027_9.png]
Figure 10
Figure 10. Figure 10: Test accuracy comparison on CIFAR10 versus SNRs with a training SNR of [PITH_FULL_IMAGE:figures/full_fig_p028_10.png]
Figure 11
Figure 11. Figure 11: Test accuracy comparison on CIFAR10 versus the number of labeled samples with [PITH_FULL_IMAGE:figures/full_fig_p029_11.png]
Figure 12
Figure 12. Figure 12: Ablation study for reconstruction pretext task: test accuracy comparison versus the [PITH_FULL_IMAGE:figures/full_fig_p030_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

58 extracted references · 52 canonical work pages

  1. [21]

    Learning task-oriented communication for edge inference: An information bottleneck approach,

    J. Shao, Y . Mao, and J. Zhang, “Learning task-oriented communication for edge inference: An information bottleneck approach,” IEEE J. Sel. Areas Commun. , vol. 40, no. 1, pp. 197–211, Jan. 2022

    work page 2022

  2. [43]

    Contrastive Learning based Semantic Communication for Wireless Image Transmission

    S. Tang et al. , “Contrastive learning based semantic communication for wireless image transmission,” 2023, arXiv:2304.09438. [Online]. Available: https://arxiv.org/abs/2304.09438

    work page Pith review arXiv 2023

  3. [1]

    Self-supervised learning enabled task-oriented semantic communication using limited labels,

    R. Gu, W. Xu, Z. Yang, X. You, and D. Niyato, “Self-supervised learning enabled task-oriented semantic communication using limited labels,” accepted by Wireless Opt. Commun. Conf. (WOCC) , 2024. 33

    work page 2024

  4. [2]

    Edge artificial intelligence for 6G: Vision, enabling technologies, and applications,

    K. B. Letaief, Y . Shi, J. Lu, and J. Lu, “Edge artificial intelligence for 6G: Vision, enabling technologies, and applications,” IEEE J. Sel. Areas Commun. , vol. 40, no. 1, pp. 5–36, Nov. 2021

    work page 2021

  5. [3]

    Toward ubiquitous and intelligent 6G networks: From architecture to technology,

    W. Xu, Y . Huang, W. Wang, F. Zhu, and X. Ji, “Toward ubiquitous and intelligent 6G networks: From architecture to technology,” Sci. China Inf. Sci. , vol. 66, no. 3, pp. 130 300:1–2, Feb. 2023

    work page 2023

  6. [4]

    Byzantine-resilient over-the-air federated learning under zero-trust architecture,

    J. Yao, W. Shi, W. Xu, Z. Yang, A. Lee Swindlehurst, and D. Niyato, “Byzantine-resilient over-the-air federated learning under zero-trust architecture,” IEEE J. Sel. Areas Commun. , pp. 1–1, Apr. 2025

    work page 2025

  7. [5]

    Combating interference for over-the-air federated learning: A statistical approach via ris,

    W. Shi, J. Yao, W. Xu, J. Xu, X. You, Y . C. Eldar, and C. Zhao, “Combating interference for over-the-air federated learning: A statistical approach via ris,” IEEE Trans. Signal Process. , vol. 73, pp. 936–953, Jan. 2025

    work page 2025

  8. [6]

    Distributed machine learning for multiuser mobile edge computing systems,

    Y . Guo et al. , “Distributed machine learning for multiuser mobile edge computing systems,” IEEE J. Sel. Top. Signal Process., vol. 16, no. 3, pp. 460–473, Apr. 2022

    work page 2022

Show all 58 references
  1. [7]

    Wireless federated learning over resource-constrained networks: Digital versus analog transmissions,

    J. Yao, W. Xu, Z. Yang, X. You, M. Bennis, and H. V . Poor, “Wireless federated learning over resource-constrained networks: Digital versus analog transmissions,” IEEE Trans. Wireless Commun. , vol. 23, no. 10, pp. 14 020–14 036, Jun. 2024

    2024

  2. [8]

    Reconfiguring wireless environments via intelligent surfaces for 6g: reflection, modulation, and security,

    J. Xu, C. Yuen, C. Huang, N. Ul Hassan, G. C. Alexandropoulos, M. Di Renzo, and M. Debbah, “Reconfiguring wireless environments via intelligent surfaces for 6g: reflection, modulation, and security,” Sci. China Inf. Sci. , vol. 66, no. 3, pp. 130 304:1–20, Feb. 2023

    2023

  3. [9]

    Beyond transmitting bits: Context, semantics, and task-oriented communications,

    D. G ¨und¨uz, Z. Qin, I. E. Aguerri, H. S. Dhillon, Z. Yang, A. Yener, K. K. Wong, and C.-B. Chae, “Beyond transmitting bits: Context, semantics, and task-oriented communications,” IEEE J. Sel. Areas Commun. , vol. 41, no. 1, pp. 5–41, Nov. 2022

    2022

  4. [10]

    Compressing semantic information with varying priorities,

    B. Guler and A. Yener, “Compressing semantic information with varying priorities,” in Proc. Data Compress. Conf. , Snowbird, UT, Mar. 2014, pp. 213–222

    2014

  5. [11]

    Semantic index assignment,

    ——, “Semantic index assignment,” in IEEE Int. Conf. Pervasive Comput. Commun. Workshops , Budapest, Hungary, Mar. 2014, pp. 431–436

    2014

  6. [12]

    Stacked intelligent metasurfaces for task-oriented semantic communications,

    G. Huang, J. An, Z. Yang, L. Gan, M. Bennis, and M. Debbah, “Stacked intelligent metasurfaces for task-oriented semantic communications,” IEEE Wireless Commun. Lett. , early access, 2024

    2024

  7. [13]

    Semantic communications: Principles and challenges,

    Z. Qin, X. Tao, J. Lu, W. Tong, and G. Y . Li, “Semantic communications: Principles and challenges,” 2022, arXiv:2201.01389. [Online]. Available: https://arxiv.org/abs/2201.01389

    2022 arXiv

  8. [14]

    Semantic text compression for classification,

    E. Kutay and A. Yener, “Semantic text compression for classification,” in Proc. IEEE Int. Conf. Commun. Workshops , Rome, Italy, May 2023, pp. 1368–1373

    2023

  9. [15]

    Classification-oriented semantic wireless communications,

    ——, “Classification-oriented semantic wireless communications,” in Proc. IEEE Int. Conf. Acoust. Speech Signal Process., Seoul, Korea, Apr. 2024, pp. 9096–9100

    2024

  10. [16]

    Deep learning-based end-to-end wireless communication systems with conditional GANs as unknown channels,

    H. Ye et al. , “Deep learning-based end-to-end wireless communication systems with conditional GANs as unknown channels,” IEEE Trans. Wireless Commun. , vol. 19, no. 5, pp. 3133–3143, May 2020

    2020

  11. [17]

    Less data, more knowledge: Building next generation semantic communication networks,

    C. Chaccour, W. Saad, M. Debbah, Z. Han, and H. V . Poor, “Less data, more knowledge: Building next generation semantic communication networks,” IEEE Commun. Surv. Tutor., early access, Jun. 2024

    2024

  12. [18]

    Semantic-preserved communication system for highly efficient speech transmission,

    T. Han, Q. Yang, Z. Shi, S. He, and Z. Zhang, “Semantic-preserved communication system for highly efficient speech transmission,” IEEE J. Sel. Areas Commun. , vol. 41, no. 1, pp. 245–259, Jan. 2023

    2023

  13. [19]

    Deep joint source-channel coding for wireless image transmission,

    E. Bourtsoulatze, D. B. Kurka, and D. G ¨und¨uz, “Deep joint source-channel coding for wireless image transmission,” IEEE Trans. Cogn. Commun. Netw. , vol. 5, no. 3, pp. 567–579, May 2019

    2019

  14. [20]

    Deep learning enabled semantic communication systems,

    H. Xie et al., “Deep learning enabled semantic communication systems,” IEEE Trans. Signal Process., vol. 69, pp. 2663– 2675, Apr. 2021. 34

    2021

  15. [22]

    Deep learning-enabled semantic communication systems with task-unaware transmitter and dynamic data,

    H. Zhang et al. , “Deep learning-enabled semantic communication systems with task-unaware transmitter and dynamic data,” IEEE J. Sel. Areas Commun. , vol. 41, no. 1, pp. 170–185, Jan. 2023

    2023

  16. [23]

    Task-oriented multi-user semantic communications,

    H. Xie et al. , “Task-oriented multi-user semantic communications,” IEEE J. Sel. Areas Commun. , vol. 40, no. 9, pp. 2584–2597, Jul. 2022

    2022

  17. [24]

    Wireless image retrieval at the edge,

    M. Jankowski et al., “Wireless image retrieval at the edge,” IEEE J. Sel. Areas Commun. , vol. 39, no. 1, pp. 89–100, Nov. 2020

    2020

  18. [25]

    Deep learning in physical layer communications,

    Z. Qin, H. Ye, G. Y . Li, and B.-H. F. Juang, “Deep learning in physical layer communications,” IEEE Wireless Commun., vol. 26, no. 2, pp. 93–99, Apr. 2019

    2019

  19. [26]

    Model-driven deep learning for physical layer communications,

    H. He, S. Jin, C.-K. Wen, F. Gao, G. Y . Li, and Z. Xu, “Model-driven deep learning for physical layer communications,” IEEE Wireless Commun., vol. 26, no. 5, pp. 77–83, Oct. 2019

    2019

  20. [27]

    Semi-supervised end-to-end learning for integrated sensing and communications,

    J. M. Mateos-Ramos et al. , “Semi-supervised end-to-end learning for integrated sensing and communications,” in IEEE Int. Conf. Mach. Learn. Commun. Netw. , Stockholm, Sweden, May 2024, pp. 132–138

    2024

  21. [28]

    Stacked intelligent metasurfaces for wireless sensing and communication: Applications and challenges,

    H. Liu, J. An, X. Jia, S. Lin, X. Yao, L. Gan, B. Clerckx, C. Yuen, M. Bennis, and M. Debbah, “Stacked intelligent metasurfaces for wireless sensing and communication: Applications and challenges,” 2024, 2407.03566. [Online]. Available: https://arxiv.org/abs/1810.00740

    2024 arXiv

  22. [29]

    A review of IoT application in a smart traffic management system,

    M. K. M. Rabby, M. M. Islam, and S. M. Imon, “A review of IoT application in a smart traffic management system,” in IEEE Int. Conf. Adv. Electr. Eng. , Dhaka, Bangladesh, Sep. 2019, pp. 280–285

    2019

  23. [30]

    A survey of intelligent network slicing management for industrial IoT: Integrated approaches for smart transportation, smart energy, and smart factory,

    Y . Wu, H.-N. Dai, H. Wang, Z. Xiong, and S. Guo, “A survey of intelligent network slicing management for industrial IoT: Integrated approaches for smart transportation, smart energy, and smart factory,” IEEE Commun. Surv. Tutor., vol. 24, no. 2, pp. 1175–1211, Mar. 2022

    2022

  24. [31]

    Enhancing performance of deep learning models with different data augmentation techniques: A survey,

    C. Khosla and B. S. Saini, “Enhancing performance of deep learning models with different data augmentation techniques: A survey,” in IEEE Int. Conf. Intell. Eng. Manag. London, UK: IEEE, Jun. 2020, pp. 79–85

    2020

  25. [32]

    Semantic data augmentation with generative models,

    C. Shivashankar and S. Miller, “Semantic data augmentation with generative models,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., Vancouver, BC, Canada, Jun. 2023, pp. 863–873

    2023

  26. [33]

    A survey on transfer learning,

    S. J. Pan and Q. Yang, “A survey on transfer learning,” IEEE Trans. Knowl. Data Eng. , vol. 22, no. 10, pp. 1345–1359, Oct. 2009

    2009

  27. [34]

    Model-agnostic meta-learning for fast adaptation of deep networks,

    C. Finn, P. Abbeel, and S. Levine, “Model-agnostic meta-learning for fast adaptation of deep networks,” in PMLR Int. Conf. Mach. Learn. Sydney, Australia: PMLR, Aug. 2017, pp. 1126–1135

    2017

  28. [35]

    Improving the generalization of adversarial training with domain adaptation,

    C. Song et al. , “Improving the generalization of adversarial training with domain adaptation,” 2018, arXiv:1810.00740. [Online]. Available: https://arxiv.org/abs/1810.00740

    2018 arXiv

  29. [36]

    Pre-trained models: Past, present and future,

    X. Han et al., “Pre-trained models: Past, present and future,” AI Open, vol. 2, pp. 225–250, Aug. 2021

    2021

  30. [37]

    Context encoders: Feature learning by inpainting,

    D. Pathak et al., “Context encoders: Feature learning by inpainting,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. , Las Vegas, NV , USA, Jun. 2016, pp. 2536–2544

    2016

  31. [38]

    Unsupervised feature learning via non-parametric instance discrimination,

    Z. Wu et al. , “Unsupervised feature learning via non-parametric instance discrimination,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., Salt Lake City, UT, USA, Dec. 2018, pp. 3733–3742

    2018

  32. [39]

    Contrastive self-supervised learning: a survey on different architectures,

    A. Khan, S. AlBarri, and M. A. Manzoor, “Contrastive self-supervised learning: a survey on different architectures,” in Proc. IEEE Int. Conf. Artif. Intell. , Islamabad, Pakistan, Mar. 2022, pp. 1–6

    2022

  33. [40]

    Self-supervised human activity recognition with localized time-frequency contrastive representation learning,

    S. R. Taghanaki et al., “Self-supervised human activity recognition with localized time-frequency contrastive representation learning,” 2022, arXiv:2209.00990. [Online]. Available: https://arxiv.org/abs/2209.00990 35

    2022 arXiv

  34. [41]

    Disentangling learnable and memorizable data via contrastive learning for semantic communications,

    C. Chaccour and W. Saad, “Disentangling learnable and memorizable data via contrastive learning for semantic communications,” 2022, arXiv:2212.09071. [Online]. Available: https://arxiv.org/abs/2212.09071

    2022 arXiv

  35. [42]

    A simple framework for contrastive learning of visual representations,

    T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A simple framework for contrastive learning of visual representations,” in Proc. IEEE Int.Conf. Mach. Learn. , Virtual Event, Jul. 2020, pp. 1597–1607

    2020

  36. [44]

    The information bottleneck method,

    N. Tishby, F. C. Pereira, and W. Bialek, “The information bottleneck method,” in Proc. Annu. Allerton Conf. Commun. Control Comput., Monticello, IL, USA, Sep. 1999, pp. 368–377

    1999

  37. [45]

    Self-supervised learning from a multi-view perspective,

    Y .-H. H. Tsai, Y . Wu, R. Salakhutdinov, and L.-P. Morency, “Self-supervised learning from a multi-view perspective,” 2020, arXiv:2006.05576. [Online]. Available: https://arxiv.org/abs/2006.05576

    2020 arXiv

  38. [46]

    Disentangled representation learning empowered CSI feedback using implicit channel reciprocity in FDD massive MIMO,

    W. Xu, J. Wu, S. Jin, X. You, and Z. Lu, “Disentangled representation learning empowered CSI feedback using implicit channel reciprocity in FDD massive MIMO,” IEEE Trans. Wireless Commun., vol. 23, no. 10, pp. 15 169–15 184, Jul. 2024

    2024

  39. [47]

    Edge learning for B5G networks with distributed signal processing: Semantic communication, edge computing, and wireless sensing,

    W. Xu, Z. Yang, D. W. K. Ng, M. Levorato, Y . C. Eldar, and M. Debbah, “Edge learning for B5G networks with distributed signal processing: Semantic communication, edge computing, and wireless sensing,” IEEE J. Sel. Top. Signal Process. , vol. 17, no. 1, pp. 9–39, Jan. 2023

    2023

  40. [48]

    DIBAD: A disentangled information bottleneck adversarial defense method using hilbert-schmidt independence criterion for spectrum security,

    S. Zhang, Y . Yang, Z. Zhou, Z. Sun, and Y . Lin, “DIBAD: A disentangled information bottleneck adversarial defense method using hilbert-schmidt independence criterion for spectrum security,” IEEE Trans. Inf. Forensics Secur., vol. 19, pp. 3879–3891, Mar. 2024

    2024

  41. [49]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., Las Vegas, NV , USA, Jun. 2016, pp. 770–778

    2016

  42. [50]

    OFDM-guided deep joint source channel coding for wireless multipath fading channels,

    M. Yang, C. Bian, and H.-S. Kim, “OFDM-guided deep joint source channel coding for wireless multipath fading channels,” IEEE Trans. Cogn. Commun. Netw. , vol. 8, no. 2, pp. 584–599, Feb. 2022

    2022

  43. [51]

    An information theoretic framework for multi-view learning,

    K. Sridharan and S. M. Kakade, “An information theoretic framework for multi-view learning,” in Proc. Annu. Conf. Learn. Theory, Helsinki, Finland, Jul. 2008, pp. 403–414

    2008

  44. [52]

    Representation learning with contrastive predictive coding,

    A. V . D. Oord, Y . Li, and O. Vinyals, “Representation learning with contrastive predictive coding,” 2018, arXiv:1807.03748. [Online]. Available: https://arxiv.org/abs/1807.03748

    2018 arXiv

  45. [53]

    Disentangled representation learning for RF fingerprint extraction under unknown channel statistics,

    R. Xie et al., “Disentangled representation learning for RF fingerprint extraction under unknown channel statistics,” IEEE Trans. Commun., vol. 71, pp. 3946–3962, Jul. 2023

    2023

  46. [54]

    Neural joint source-channel coding,

    K. Choi et al. , “Neural joint source-channel coding,” in Proc. Int.Conf. Mach. Learn. Long Beach, CA, USA: PMLR, Jun. 2019, pp. 1182–1192

    2019

  47. [55]

    Learning multiple layers of features from tiny images,

    A. Krizhevsky, G. Hinton et al., “Learning multiple layers of features from tiny images,” Tech. Rep., 2009

    2009

  48. [56]

    Reading digits in natural images with unsupervised feature learning,

    Y . Netzer et al., “Reading digits in natural images with unsupervised feature learning,” in Proc. Annu. Conf. Neural Inf. Process. Syst. Workshop, Granada, Spain, Dec. 2011, pp. 12–17

    2011

  49. [57]

    Flowers recognition dataset,

    A. Mamaev, “Flowers recognition dataset,” 2021. [Online]. Available: https://www.kaggle.com/datasets/alxmamaev/ flowers-recognition

    2021

  50. [58]

    Seq2seq-sc: End-to-end semantic communication systems with pre-trained language model,

    J.-H. Lee et al. , “Seq2seq-sc: End-to-end semantic communication systems with pre-trained language model,” 2022, arXiv:2210.15237. [Online]. Available: https://arxiv.org/abs/2210.15237

    2022 arXiv

Pith tools

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