Pith. sign in

REVIEW 3 major objections 10 minor 1 cited by

Adaptive Sampling and Joint Semantic-Channel Coding under Dynamic Channel Environment

T0 review · 3 major / 10 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read This paper claims that adding semantic-aware adaptive sampling to joint source-channel coding can cut the number of measurements an image transmitter must take while preserving or improving reconstruction quality, all with a single model…

desk verdict A plausible integration of semantic sampling and SNR-conditioned JSCC, but the headline gains depend on a lossless, cost-free side map the paper never charges for. read the letter →

arxiv 2502.07236 v1 pith:SJEM7DCH submitted 2025-02-11 eess.IV

Zhiyuan Qi , Yulong Feng , Zhijin Qin This is my paper · ORCID
classification eess.IV
keywords semanticcommunicationadaptivesamplingcompressedsensingjointsource-channelcodingimagetransmissionchannelSNRadaptationattentionmoduledeeplearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that the data-acquisition step, not just source-channel coding, belongs inside an end-to-end semantic communication system. It introduces Adaptive-JSSCC, which measures the semantic importance of each image block, allocates a per-block sampling ratio through a distribution map $R$, and uses that same map at the receiver to guide reconstruction. The claimed payoff is that fewer measurements are needed for a given reconstruction quality than with uniform sampling, and that a single model trained over an SNR range works at varying channel conditions where a fixed-SNR model would fail. The reported simulations support a roughly $0.5$ to $1$ dB PSNR gain from semantic over uniform sampling, and memory savings versus storing multiple SNR-specific models.

What carries the argument

The semantic sampling ratio distribution map $R$ is the load-bearing object: it is computed at the transmitter from the image and SNR $\mu$, it is shared with the receiver (assumed losslessly), and it controls both compressed-sensing sampling matrices per block and the proximal-gradient reconstruction iterations. The second mechanism is the attention-based channel adaptive module (ACAM), which concatenates the current SNR with pooled feature statistics, passes them through a small network, and multiplies the resulting channel attention weights into the features at sampling, encoding, decoding, and reconstruction stages. Together they let a single model allocate samples according to semantic importance and adapt its representations when the channel SNR changes.

What would settle it

Quantize $R$ to a realistic number of bits, transmit it over the same AWGN channel, and measure total-rate-matched PSNR against uniform sampling with no map. If the adaptive system no longer beats uniform sampling at equal total rate, the claimed advantage disappears.

Watch

Extended reading notes

Core claim

The central claim is that an image transmission system can jointly optimize how many samples to take, where to take them, and how to encode them for a noisy channel. Semantic saliency, evaluated by a scanning network from the image and the current SNR $\mu$, determines a sampling ratio for every block; these ratios form the map $R$ that is shared with the receiver. The encoder and decoder are augmented with attention-based channel adaptive modules that take $\mu$ as an extra input, so the same trained model can rescale its features for different channel conditions instead of being retrained per SNR. With this design, the paper shows that semantic sampling matches or beats uniform sampling at lower sampling ratios, and the adaptive system outperforms JSSCC models trained at a single SNR when the test SNR moves away from the training value.

Load-bearing premise

The load-bearing premise is that the semantic sampling ratio distribution map $R$ is transmitted from the transmitter to the receiver without any loss and without consuming rate; the paper states this assumption explicitly, and the reconstruction algorithm depends on $R$ at every iteration.

Editorial extensions

If this is right

  • A transmitter can spend fewer measurements on flat or unimportant regions without sacrificing image quality, which lowers sensing energy and data-acquisition cost.
  • A single Adaptive-JSSCC model, trained with SNR drawn uniformly from $[0,20]$ dB, can serve a varying channel where previously several JSSCC models trained at fixed SNRs had to be stored and switched.
  • The semantic map $R$ makes reconstruction depend on knowing where samples were taken; the method's gains are contingent on that side information being available.
  • At low channel bandwidth ratios, performance saturates earlier, so adaptive sampling helps most in the sample-limited rather than bandwidth-limited regime.

Reading between the lines

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

  • If $R$ has to be quantized and sent over the same channel, the reported $0.5$-$1$ dB gain must be weighed against the rate those bits consume; at sufficiently low total rate, uniform sampling could close or reverse the gap.
  • The zero-loss $R$ assumption also implies the receiver knows the exact sampling pattern; under fast fading or feedback delay, a stale $R$ would misguide reconstruction, so the method's robustness claim likely degrades when channel dynamics outpace map delivery.
  • The same per-block allocation idea could be applied to video frames or to task-oriented metrics such as detection or segmentation instead of PSNR, with the semantic scanning network retrained for those objectives.
  • An obvious stress test is to vary the number of reconstruction iterations and the block size: the reported gains may trade off against compute and edge artifacts.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 10 minor

Summary. This paper proposes Adaptive-JSSCC, an end-to-end trainable system for image transmission over AWGN channels that couples content-adaptive sampling with joint source-channel coding. At the transmitter, a channel-conditioned saliency network assigns a per-block sampling ratio for B by B blocks, producing a sampling-ratio distribution map R that is shared with the receiver; each block is measured with its own learned sampling matrix, encoded by an SNR-adaptive encoder, and sent over the channel. At the receiver, the decoder output is refined by Np = 11 unrolled proximal gradient descent iterations whose proximal mapping is conditioned on R, and an attention-based channel adaptive module (ACAM) scales features in the sampler, encoder, decoder, and reconstructor according to the instantaneous SNR. Experiments compare the full system with the authors' own uniform-sampling and SNR-fixed ablations on Set11, Kodak24, and BSD68 over sampling ratios from 0.01 to 0.50 and channel bandwidth ratios of 1, 1/2, and 1/4, reporting PSNR/SSIM, together with a model-storage comparison claiming that one Adaptive-JSSCC model replaces five SNR-specific JSSCC models.

Significance. If the results hold, the paper's significance lies in combining two lines of work usually treated separately, namely content-aware acquisition and channel-robust learned coding, into a single trainable framework, with a concrete mechanism (ACAM) for adapting both sampling and coding to a time-varying SNR. Strengths of the manuscript are its falsifiable quantitative predictions (Tables II-III and Figs. 4-6), the wide sampling-ratio grid down to r = 0.01, the three public test sets, and the explicit accounting of model storage, a practical issue often ignored in this literature. The main weaknesses are that the semantic-versus-uniform comparison assumes lossless, zero-rate delivery of the side information R (a limitation stated in Sec. III-B), the baseline set contains no external state-of-the-art system, and all reported numbers come from a single run with one PSNR inversion in Table III. The contribution is plausible and the framework is well motivated, but the evidence as presented does not yet establish the headline acquisition-reduction and state-of-the-art claims.

major comments (3)
  1. [Sec. III-B; Eq. (11); Algorithm 2] The central claim that semantic sampling reduces data acquisition without degrading reconstruction is evaluated under a stated idealization that is load-bearing for the comparison. The manuscript states in Sec. III-B that the transmission of R proceeds without any loss, and Eq. (11) and Algorithm 2 condition every unrolled proximal-gradient iteration on R, whereas the uniform baseline Adaptive-JSCC requires no such side information; the comparison is therefore not made at equal end-to-end rate. The overhead is material in the low-rate regime that motivates the paper: for a 128 by 128 image with B = 32 and r = 0.01 there are only about 164 sampled measurements, while R carries 16 per-block ratios that must be quantized and delivered exactly, adding roughly 2-20% of the payload depending on the quantization of measurements and ratios, and this ignores the error-protection coding needed to make R lossless over the same AWGN channel. Because the decoder conditions all 11 iterations on R, any quantization or corruption of R changes the reconstructed image in a way the paper does not measure. Please report the bit budget of R, include it in the rate or CBR comparison against uniform sampling, and add experiments with quantized or corrupted R; also clarify that the saving is in transmitted measurements rather than acquired pixels, since the saliency map is computed from the full image at the transmitter.
  2. [Table III; Sec. IV-B(b)] The quantitative claim that semantic sampling typically yields an average PSNR improvement of around 0.5 to 1 dB is not fully supported by Table III. Averaging the row-wise PSNR differences across the six sampling ratios gives roughly 0.42 dB on Set11, 0.40 dB on Kodak24, and 0.44 dB on BSD68, so the typical gain is closer to 0.4 dB, and one cell is inverted: on Set11 at r = 0.04 the adaptive method is 26.16 dB against 26.65 dB for uniform sampling, while the SSIM in the same cell favors the semantic method (0.8087 vs 0.7980). The inversion and the PSNR/SSIM disagreement need an explanation, and the absence of error bars or multi-seed statistics means the gains cannot be distinguished from seed-dependent variation; the SSIM gains are positive in all 18 cells, which is encouraging, but the PSNR claim should be restated with the actual mean and variance over repeated runs.
  3. [Abstract; Sec. IV, Table I] The abstract claims an improvement in comparison to the state of the art, but the experimental section compares only against the authors' own ablations. Adaptive-JSCC (uniform sampling), JSSCC, and JSCC are defined in Table I as variants of the proposed architecture, and the JSSCC-n baselines in Table II are simply n copies of the same model trained at different SNRs; no external state-of-the-art system such as DeepJSCC [4], DeepJSCC-Q [13], or a recent learned compressed-sensing method is evaluated under the same conditions. The claim should be re-scoped to comparisons against the authors' own ablations, or external baselines should be added; as it stands, the state-of-the-art wording in the abstract is not supported by the evidence in the paper.
minor comments (10)
  1. [Eq. (8)] Equation (8) as printed has an unbalanced parenthesis in the expression for kappa; please check the formula.
  2. [Sec. III-C] The semantic extracting network is introduced as 'mathcalE', an unrendered LaTeX command, and is then used as E(k) in Eq. (11) and Algorithm 2; please unify the notation and define its inputs and outputs.
  3. [Algorithm 1] Line 4 of Algorithm 1, 'ri <- r, M;', does not specify the functional relationship that maps the overall ratio r and saliency map M to the per-block ratio ri; an explicit formula is needed for reproducibility.
  4. [Sec. IV-A(b)] The sentence 'q is randomly equivocally selected' presumably means q is selected uniformly; as printed it is ambiguous.
  5. [Table I] The column header 'Sampling -Reconstruction(w/o)' is ambiguous; please specify whether a checkmark denotes the presence or the absence of semantic sampling-reconstruction for each ablation model.
  6. [Table II] Please state how the 'better average PSNR' of the multi-model baselines is computed (for example, taking the best model at each test SNR and then averaging), because the memory-versus-performance comparison depends on this rule.
  7. [Fig. 6] Please state whether the CBR is fixed at training time or adjustable at test time for each curve, and support the assertion that performance achieves saturation earlier for models trained at lower CBR with the corresponding data.
  8. [Abstract] The abstract contains an editing artifact in 'we optimize sampling matrices for each region of the most individually'; please rephrase.
  9. [Fig. 4] The channel-adaptation experiments cover only the training SNR range of 0 to 20 dB; tests at out-of-range SNRs would better substantiate the claimed robustness to dynamic channel environments.
  10. [General] No code or trained models are released; given the sensitivity of the unrolled PGD reconstruction to Np and to the learned sampling policy, releasing the implementation would materially aid reproducibility.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity: the PSNR/SSIM results are empirical measurements, SNR is a conditioning input rather than a fitted constant, and the self-citations are contextual only.

full rationale

The central claim that Adaptive-JSSCC reduces data acquisition without degrading reconstruction is supported by measured PSNR/SSIM comparisons in Table III and Fig. 6, not by a derivation that reduces to its own inputs. The semantic sampling ratio distribution map R is produced by a learned network and the receiver reconstruction is conditioned on it in Eq. (11); this is an algorithmic design choice rather than a definitional circularity. The paper explicitly assumes that the transmission of R is without any loss (Section III-B), which is an overhead and fairness caveat for the comparison against uniform sampling, but it is not a circular step because the reported gains are experimental measurements under that stated assumption. SNR is used as an extra conditioning input to the attention module, not as a fitted constant that is later renamed as a prediction. The only self-citations are background references for semantic communication and related sampling work ([2], [3], [5], [7]); they do not carry the load of the main result. No equation in the paper was found to be equivalent to its input by construction, and no fitted parameter was presented as an independent prediction. Therefore the appropriate finding is minor self-citation without load-bearing circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on standard deep-network approximation plus a few hand-set hyperparameters. No new physical entities, forces, or conserved quantities are introduced. The most expensive unstated assumption is the lossless, cost-free transmission of the semantic sampling map R; the objective in Eq. (6) does not include any rate or distortion term for R, and every reported gain assumes it arrives perfectly at the decoder.

free parameters (4)
  • Block size B = 32
    Chosen for sampling and reconstruction; controls the number of blocks l=16 and the granularity of semantic importance allocation.
  • Reconstruction iterations Np = 11
    Number of unrolled proximal gradient iterations, set by hand; the paper does not provide a convergence or sensitivity study.
  • Training SNR range = [0,20] dB uniformly sampled
    The adaptation claim is only demonstrated within this training SNR range; results outside it are not shown.
  • Sampling ratio training distribution = q uniform in [1,N] per block
    The overall sampling ratio r is drawn uniformly during training; performance at very low ratios depends on this choice.
assumptions (5)
  • domain assumption The channel is AWGN with known SNR mu at both transmitter and receiver (Eq. 3).
    All training and testing use this model; ACAM and the semantic scanning network take mu as an input, so accurate SNR knowledge at both ends is assumed.
  • domain assumption The semantic sampling ratio distribution map R is received without loss or rate cost (Section III-B).
    Reconstruction depends on R; the paper explicitly assumes lossless transmission of R and does not account for its bandwidth or error sensitivity in the objective.
  • domain assumption Block-wise compressed sensing with separately assigned sampling matrices is a valid model of image acquisition.
    The system divides images into B by B blocks and samples each with its own matrix; no conditioning or block-boundary analysis is given.
  • domain assumption A learned semantic scanning network can predict saliency from image and SNR in a way that transfers to unseen test sets.
    The semantic map M = S(S, mu) is learned end-to-end; its generalization is only indirectly measured through final reconstruction quality.
  • ad hoc to paper Unrolled proximal gradient descent with a learned proximal mapping reaches a good reconstruction in Np = 11 iterations.
    The number of unrolled iterations is fixed by hand without a convergence analysis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Adaptive Sampling and Joint Semantic-Channel Coding under Dynamic Channel Environment." pith.science (2026). https://pith.science/paper/SJEM7DCH

@misc{pith2026250207236,
  author       = {Pith},
  title        = {Pith review of: Adaptive Sampling and Joint Semantic-Channel Coding under Dynamic Channel Environment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SJEM7DCH}},
  note         = {Machine review of arXiv:2502.07236}
}
read the original abstract

Deep learning enabled semantic communications are attracting extensive attention. However, most works normally ignore the data acquisition process and suffer from robustness issues under dynamic channel environment. In this paper, we propose an adaptive joint sampling-semantic-channel coding (Adaptive-JSSCC) framework. Specifically, we propose a semantic-aware sampling and reconstruction method to optimize the number of samples dynamically for each region of the images. According to semantic significance, we optimize sampling matrices for each region of the most individually and obtain a semantic sampling ratio distribution map shared with the receiver. Through the guidance of the map, high-quality reconstruction is achieved. Meanwhile, attention-based channel adaptive module (ACAM) is designed to overcome the neural network model mismatch between the training and testing channel environment during sampling-reconstruction and encoding-decoding. To this end, signal-to-noise ratio (SNR) is employed as an extra parameter input to integrate and reorganize intermediate characteristics. Simulation results show that the proposed Adaptive-JSSCC effectively reduces the amount of data acquisition without degrading the reconstruction performance in comparison to the state-of-the-art, and it is highly adaptable and adjustable to dynamic channel environments.

Figures

Figures reproduced from arXiv: 2502.07236 by the authors.

Figure 1
Figure 1. The framework of the proposed Adaptive-JSSCC, which contains semantic sampling, joint semantic-channel codec and [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The proposed joint semantic-channel codec. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The architecture of semantic scanning network. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Details with uniform (left) and semantic (right) method [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Performance comparison between semantic and uniform [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. AirLLM: Diffusion Policy-based Adaptive LoRA for Remote Fine-Tuning of LLM over the Air

    cs.LG 2025-07 conditional novelty 6.0 of 10

    A PPO-plus-diffusion policy adaptively allocates LoRA ranks per layer based on channel SNR and data complexity, improving accuracy by up to 0.69% and cutting transmitted parameters by 12.5% over AdaLoRA.

Reference graph

Works this paper leans on

19 extracted references · 16 canonical work pages · cited by 1 Pith paper

  1. [4]

    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 Transactions on Cognitive Communications and Networking , vol. 5, no. 3, pp. 567–579, 2019

    work page 2019

  2. [13]

    DeepJSCC-Q: Constellation constrained deep joint source-channel coding,

    T. Y . Tung, D. B. Kurka, M. Jankowski, and D. G ¨und¨uz, “DeepJSCC-Q: Constellation constrained deep joint source-channel coding,” IEEE Journal on Selected Areas in Information Theory, vol. 3, no. 4, pp. 720–731, 2022

    work page 2022

  3. [1]

    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 Journal on Selected Areas in Communications, vol. 41, no. 1, pp. 5–41, 2022

    2022

  4. [2]

    AI em- powered wireless communications: From bits to semantics,

    Z. Qin, L. Liang, Z. Wang, S. Jin, X. Tao, W. Tong, and G. Y . Li, “AI em- powered wireless communications: From bits to semantics,” Proceedings of the IEEE , 2024

    work page 2024

  5. [3]

    Semantic communications: Principles and challenges,

    Z. Qin, X. Tao, J. Lu, W. Tong, and G. Y . Li, “Semantic communications: Principles and challenges,” arXiv preprint arXiv:2201.01389 , 2021

    arXiv 2021

  6. [5]

    Semantic-aware image compressed sensing,

    B. Zhang, Z. Qin, and G. Y . Li, “Semantic-aware image compressed sensing,” in 2023 IEEE 33rd International Workshop on Machine Learning for Signal Processing (MLSP) . IEEE, 2023, pp. 1–6

    work page 2023

  7. [6]

    Variable-rate deep image compres- sion through spatially-adaptive feature transform,

    M. Song, J. Choi, and B. Han, “Variable-rate deep image compres- sion through spatially-adaptive feature transform,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 2380– 2389

    work page 2021

  8. [7]

    Compression ratio learning and semantic communications for video imaging,

    B. Zhang, Z. Qin, and G. Y . Li, “Compression ratio learning and semantic communications for video imaging,” IEEE Journal of Selected Topics in Signal Processing, 2024

    work page 2024

Show all 19 references
  1. [8]

    Deep coded aperture design: An end-to-end approach for computational imaging tasks,

    J. Bacca, T. Gelvez-Barrera, and H. Arguello, “Deep coded aperture design: An end-to-end approach for computational imaging tasks,” IEEE Transactions on Computational Imaging , vol. 7, pp. 1148–1160, 2021

    2021

  2. [9]

    Time-multiplexed coded aperture imaging: Learned coded aperture and pixel exposures for compressive imaging systems,

    E. Vargas, J. N. Martel, G. Wetzstein, and H. Arguello, “Time-multiplexed coded aperture imaging: Learned coded aperture and pixel exposures for compressive imaging systems,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2021, pp. 2692–2702

    2021

  3. [10]

    Compressive light field photography using overcomplete dictionaries and optimized projec- tions,

    K. Marwah, G. Wetzstein, Y . Bando, and R. Raskar, “Compressive light field photography using overcomplete dictionaries and optimized projec- tions,” ACM Transactions on Graphics (TOG) , vol. 32, no. 4, pp. 1–12, 2013

    2013

  4. [11]

    Wireless image transmission using deep source channel coding with attention modules,

    J. Xu, B. Ai, W. Chen, A. Yang, P. Sun, and M. Rodrigues, “Wireless image transmission using deep source channel coding with attention modules,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, no. 4, pp. 2315–2328, 2021

    2021

  5. [12]

    Predic- tive and adaptive deep coding for wireless image transmission in semantic communication,

    W. Zhang, H. Zhang, H. Ma, H. Shao, N. Wang, and V . C. Leung, “Predic- tive and adaptive deep coding for wireless image transmission in semantic communication,” IEEE Transactions on Wireless Communications, vol. 22, no. 8, pp. 5486–5501, 2023

    2023

  6. [14]

    Content-aware scalable deep compressed sensing,

    B. Chen and J. Zhang, “Content-aware scalable deep compressed sensing,” IEEE Transactions on Image Processing , vol. 31, pp. 5412–5426, 2022

    2022

  7. [15]

    Learning a deep convolutional network for image super-resolution,

    C. Dong, C. C. Loy, K. He, and X. Tang, “Learning a deep convolutional network for image super-resolution,” in Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part IV 13 . Springer, 2014, pp. 184–199

    2014

  8. [16]

    Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising,

    K. Zhang, W. Zuo, Y . Chen, D. Meng, and L. Zhang, “Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising,” IEEE Transactions on Image Processing , vol. 26, no. 7, pp. 3142–3155, 2017

    2017

  9. [17]

    Reconnet: Non-iterative reconstruction of images from compressively sensed mea- surements,

    K. Kulkarni, S. Lohit, P. Turaga, R. Kerviche, and A. Ashok, “Reconnet: Non-iterative reconstruction of images from compressively sensed mea- surements,” in Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, 2016, pp. 449–458

    2016

  10. [18]

    Beyond the kodak image set: A new reference set of color image sequences,

    S. Andriani, H. Brendel, T. Seybold, and J. Goldstone, “Beyond the kodak image set: A new reference set of color image sequences,” in 2013 IEEE International Conference on Image Processing . IEEE, 2013, pp. 2289– 2293

    2013

  11. [19]

    A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics,

    D. Martin, C. Fowlkes, D. Tal, and J. Malik, “A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics,” in Proceedings eighth IEEE International Conference on Computer Vision. ICCV 2001 , vol. 2...

    2001

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.