{"id":"12f9c092-5aca-4e24-93dd-5679d7bdd8c1","arxiv_id":"2501.04732","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A lightweight channel-adaptive Transformer for wireless image transmission that reaches higher PSNR and perceptual quality than SwinJSCC using far fewer channel-adaptation parameters and FLOPs.","lead":"SNR-EQ-JSCC is a new Transformer-based system that embeds the signal-to-noise ratio directly into the attention layers of a joint source-channel coder for wireless image transmission, and it also adjusts the attention queries according to the current channel SNR. The approach is much lighter than existing channel adaptation methods while matching or slightly improving reconstruction quality, which matters for battery-constrained edge devices.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline overhead claim contradicts Table III: ratios are 6.38% storage and 0.045% compute, not 0.05% and 6.38%.","rationale":"The Reader's weakest assumption was that the small PSNR/MS-SSIM/LPIPS gains come from a single-run evaluation without confidence intervals. That is a legitimate reproducibility concern and is part of why the manuscript should not be accepted as-is. In my reading, however, there is a more decisive, text-internal problem: the paper's headline numbers for the 'lightweight' contribution are reversed relative to its own Table III. Computing the ratios from the table gives 6.38% for parameters and 0.045% for FLOPs, which is the opposite of the Abstract's '0.05% storage, 6.38% computational complexity.' Thus one of the two central quantitative claims is demonstrably misstated, independent of stochastic variation. I therefore disagree that the single-run issue is the weakest point; the overhead claim is currently the least secure because it is contradicted by the authors' own evidence. The corrected overhead figures are still small, so the right remedy is a conditional acceptance requiring corrected text and, ideally, released training configuration and code to verify the performance margins. The verdict remains CONDITIONAL, hence UNCHANGED relative to the Reader.","tokens_in":9367,"tokens_out":8836,"duration_ms":78042,"concrete_test":"Recompute the overhead ratios from Table III using the authors' own columns: R_params = (6.29e5 + 2.42e2) / 9.86e6 and R_FLOPs = (3.93e5 + 2.03e2) / 8.72e8. If R_params is approximately 6.38% and R_FLOPs is approximately 0.045%, the Abstract and Section V must be revised to state the corrected percentages, or the table must be amended to match the text. The manuscript should also define which column counts as 'storage overhead' (parameters) and which as 'computational complexity' (FLOPs), since the current text reverses them. This single arithmetic check settles whether the headline efficiency claim is internally consistent.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The Abstract and Conclusion state that SNR-EQ-JSCC requires 0.05% of the storage overhead and 6.38% of the computational complexity of SwinJSCC's Channel ModNet for channel adaptation. The paper's own Table III contradicts this. Summing the SNR Embedding and CAQ rows gives 6.29242e5 parameters and 3.93203e5 FLOPs; the Channel ModNet row gives 9.86e6 parameters and 8.72e8 FLOPs. The resulting ratios are 6.38% for parameters (storage) and 0.0451% for FLOPs (computational complexity), i.e., the two percentages in the text are exactly inverted. The storage overhead is about 128 times larger than claimed, while the computational-complexity figure is about 141 times smaller than the claimed value. The Conclusion's statement that CAQ costs 'one hundred-thousandth and one ten-thousandth' of the Channel ModNet overhead is also not derivable from the table (the actual CAQ fractions are about 1/40,000 in parameters and 1/4,300,000 in FLOPs). This is an internal inconsistency in one of the two headline contributions, not a matter of statistical noise. The corrected numbers are still small, so a revision can fix the claim, but as written the central efficiency assertion is not supported.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SNR-EQ-JSCC, a Transformer-based joint source-channel coding scheme for wireless image transmission. The two new components are an SNR embedding added to the input of each channel-adaptive multi-head attention block and a channel-adaptive query (CAQ) module that applies a positive scaling and an offset to the queries as a function of the SNR; a Pearson-correlation penalty is added to the loss to stabilize training. For imperfect SNR feedback, the authors propose substituting the average SNR without retraining. On DIV2K, the paper reports PSNR/MS-SSIM/LPIPS gains over SwinJSCC and ADJSCC and claims extremely small channel-adaptation overhead.","tokens_in":9673,"tokens_out":10049,"duration_ms":89050,"significance":"If the empirical claims hold, the contribution is a useful, low-complexity way to make Transformer-based JSCC channel-adaptive. The design is simple and transferable, and the ablation study with 'w/o CAQ' and 'w/o EM' variants is informative. The average-SNR fallback addresses a practical feedback limitation. However, the evidence currently falls short: the reported efficiency numbers are internally inconsistent, and the performance gains rest on a single training run without statistical characterization. The corrections needed are local, so the underlying idea appears salvageable.","major_comments":[{"comment":"Section IV.F, Table III, and the Abstract/Conclusion report that the storage overhead and computational complexity of SNR-EQ-JSCC's CA module are 0.05% and 6.38% of those of SwinJSCC's Channel ModNet. The table's own numbers give the opposite assignment: (6.29e5 + 2.42e2)/9.86e6 = 6.38% for parameters and (3.93e5 + 2.03e2)/8.72e8 = 0.045% for FLOPs (which rounds to 0.05%). Thus the statement is internally inconsistent; the two percentages have been transposed. The Conclusion's claim that CAQ alone requires 'one hundred-thousandth and one ten-thousandth' of Channel ModNet is also not supported by the table, which yields 242/9.86e6 ≈ 1/41,000 parameters and 203/8.72e8 ≈ 1/4.3e6 FLOPs. Because the low-overhead claim is a headline contribution, this must be corrected and the definitions (parameters for storage, FLOPs for computation) stated consistently.","section":"Abstract, Section IV.F, Table III, Conclusion"},{"comment":"Section IV.A and Figures 2 and 3 present all performance curves as single runs with no error bars, and the text does not report the number of random seeds, optimizer, learning rate, batch size, number of epochs, or validation-based model selection. The headline gains are small (e.g., 0.32 dB PSNR and 0.0095 LPIPS at r=1/32, µtest=20 dB), so without a measure of run-to-run variation the claim that SNR-EQ-JSCC consistently outperforms SwinJSCC and ADJSCC is not yet established. Please report multi-seed means and confidence intervals (or standard deviations) and complete training details; if this is infeasible, the quantitative claims should be appropriately qualified.","section":"Section IV.A and Figures 2-3"},{"comment":"Section IV.A states that SNR-EQ-JSCC uses the SwinJSCC backbone 'but the attention block near the channel is modified to align with the divided z.' This means the comparison in Section IV.C is not a pure comparison of CA mechanisms: the backbone itself differs from SwinJSCC. The text should quantify the effect of this modification by explicitly comparing 'SNR-EQ-JSCC w/o CA' with 'SwinJSCC w/o CA'; those curves appear in Figure 3 but no numbers or discussion are given. Without this, part of the reported gain may be attributable to the structural change rather than to CAQ or SNR embedding.","section":"Section IV.A, Section IV.C"}],"minor_comments":[{"comment":"Eq. (4): the definition of softmax as e^{(·)_{i,j}} / \\sum_i \\sum_j e^{(·)_{i,j}} normalizes over all entries of the attention matrix, whereas standard scaled dot-product attention uses a row-wise softmax over keys for each query. This affects the definition of attention scores in Sections III-A and III-B and should be corrected or clarified.","section":"Eq. (4)"},{"comment":"Section II: 'the receiver feedback the fading coefficient' should read 'feeds back'; in Section IV.F, 'Channel ModeNet' should be 'Channel ModNet' for consistency with Table III.","section":"Section II and Section IV.F"},{"comment":"Section IV.A: the evaluation protocol for DIV2K is underspecified; it should state whether the 100 test images are scored as whole images or as 256×256 patches, how overlapping patches are aggregated if any, and what image size is used for the reported curves.","section":"Section IV.A"},{"comment":"Table III: the layout is difficult to parse because method names appear in the header row and 'FLOPs'/'Parameters' as row labels; use separate columns per method or a clearer multi-row layout, and state the input size and FLOPs definition used for the counts.","section":"Table III"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the headline efficiency claim is an arithmetical error that is straightforward to fix, and the requested experimental supplements should be feasible within a revision. The novelty relative to SwinJSCC and ADJSCC is incremental but adequate for a specialized journal if the empirical case is tightened. No concerns about attribution beyond the issues above."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a plausible incremental contribution to JSCC channel adaptation, but the authors manage to invert their own headline efficiency numbers in the text, and the actual gains rest on single-run evaluations.\n\nWhat's new: CAQ applies an affine transform to attention queries conditioned on SNR, plus a learned SNR embedding added to the inputs of the MHA block. Both are familiar techniques (FiLM-like conditioning, learned embeddings), but the combination in this setting, plus the correlation-based penalty terms to stabilize training, is new. The ablation study with each component removed helps the reader see what works. The alternative of running on average SNR under imperfect feedback without retraining is a nice practical add-on.\n\nWhat the paper does well: comparisons against SwinJSCC and ADJSCC on DIV2K, with ablations for lambda, CAQ, and embedding. The reported gains at higher SNR are consistent with the design. The overhead table is there, which is how we can check the claims.\n\nThe soft spots, in order of weight. First, the headline claim is internally contradicted by Table III. The text says 0.05% storage and 6.38% computational complexity for CA relative to Channel ModNet. Summing the SNR Embedding and CAQ rows gives 6.29e5 parameters and 3.93e5 FLOPs. Against ModNet's 9.86e6 parameters and 8.72e8 FLOPs, that's 6.38% of the parameters and 0.045% of the FLOPs. The two percentages are swapped. The Conclusion's \"one hundred-thousandth and one ten-thousandth\" for CAQ alone is also wrong by orders of magnitude. The corrected numbers are still small, so this is fixable in revision, but as written the central efficiency assertion is not supported by the paper's own data. Second, no error bars, no multi-seed runs, and the code isn't released. Gains of 0.1-0.3 dB on a single run are within the noise of typical training variance. Third, the penalty weight lambda is essential; without it, CAQ hurts perception metrics. That's an honest finding, but it means the method has a hand-tuned main knob. Fourth, the \"significant improvements in perception metrics\" claim is based on small absolute differences (e.g., 0.0095 LPIPS) with no variance.\n\nThe math and citation pattern look fine; no self-citation inflation, no circularity. The paper is honest about the imperfect-feedback regime.\n\nBottom line: this is a serious but incremental systems paper. It deserves a careful referee, not a desk reject, but the revision must fix the overhead numbers and add robustness evidence. Recommend: send to peer review, with the requirement that the authors reconcile Table III with the text and report multi-run statistics.","headline":"A plausible lightweight channel-adaptive JSCC scheme, but the paper's own Table III contradicts the headline efficiency claim and the gains rest on single-run evaluations.","tokens_in":10125,"tokens_out":2867,"would_cite":false,"duration_ms":25381,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A Transformer JSCC coder that conditions each attention block on channel SNR outperforms SwinJSCC with 0.05% of its channel-adaptation storage overhead.","keywords":["joint source-channel coding","semantic communication","channel adaptation","multi-head attention","Transformer","SNR embedding","channel-adaptive query","wireless image transmission"],"falsifier":"Train SNR-EQ-JSCC and SwinJSCC on the same DIV2K split with several random seeds and compare the PSNR gap at $\\mu = 20$ dB and $r = 1/32$; if the gap over seeds is not consistently positive, or if the seed-to-seed spread exceeds the 0.32 dB gain, the claim of consistent outperformance is not established.","tokens_in":9175,"feed_emoji":"📡","tokens_out":8231,"duration_ms":64244,"temperature":0.7,"pith_summary":"The paper proposes SNR-EQ-JSCC, a Transformer-based joint source-channel coding scheme whose attention blocks are conditioned directly on the channel signal-to-noise ratio $\\mu$. It claims that embedding $\\mu$ into each attention block and letting $\\mu$ rescale and shift the attention queries yields better image reconstruction than the SwinJSCC baseline across a wide SNR range, while the extra storage and computation needed for channel adaptation drop to $0.05\\%$ and $6.38\\%$ of SwinJSCC's channel-adaptation module. The paper also claims that when instantaneous SNR feedback is imperfect, feeding only the average SNR $\\bar{\\mu}$ into the same trained network still outperforms both SwinJSCC and ADJSCC. A reader would care because channel adaptation in semantic communication is usually bought with large dedicated networks; here it is folded into the attention computation itself.","feed_headline":"SNR-aware attention beats SwinJSCC with 0.05% of the adaptation cost","feed_subtitle":"Transformer coding embeds SNR into attention blocks, beating SwinJSCC on PSNR and perceptual quality.","key_machinery":"The load-bearing object is the channel-adaptive multi-head attention (CAMHA) block, which replaces the standard attention block in a Swin-Transformer-based JSCC backbone. It carries channel adaptation through two operations: a channel-adaptive query (CAQ) that rescales and shifts the queries with $q' = a_\\mu q + b_\\mu$, and an SNR embedding that adds a learned function of SNR to the block's input, $x' = x + g_{\\theta_{\\mathrm{SNR}}}(\\mu)$. The CAQ directly reshapes attention scores under channel conditions, while the SNR embedding adds a global conditioning signal; the ReLU-activated $a_\\mu$ prevents the score ranking from flipping, and the correlation-based penalty terms in the loss keep $a_\\mu$ and $b_\\mu$ tied to $\\mu$ during training. The paper also uses a two-level encoding structure: the whole image is encoded once with the average SNR, then per-fading-block portions are encoded with instantaneous SNR, so fast fading is handled without re-encoding the full image.","core_discovery":"The central claim is that a multi-head attention block can carry the entire channel-adaptation burden of a JSCC image coder by adjusting its queries according to $\\mu$, with $\\mu$ also added to the block's inputs like a learnable embedding. Concretely, each attention head computes $q' = a_\\mu q + b_\\mu$ from the SNR through small networks, with $a_\\mu > 0$ to keep attention-score rankings stable, and the block input is augmented as $x'_{\\mathrm{model}} = x_{\\mathrm{model}} + g_{\\theta_{\\mathrm{SNR}}}(\\mu)$. Penalty terms $L_a = \\mathrm{ReLU}(\\mathrm{Corr}(\\mu, a_\\mu))$ and $L_b = \\mathrm{ReLU}(\\mathrm{Corr}(\\mu, b_\\mu))$ push the query parameters to be negatively correlated with $\\mu$, which the authors argue makes the coder suppress non-essential elements at low SNR. On DIV2K image transmission, the paper reports that SNR-EQ-JSCC outperforms SwinJSCC in PSNR, MS-SSIM, and LPIPS at both compression rates tested; for example, at $r=1/32$ and $\\mu=20$ dB it gains $0.32$ dB PSNR, $0.0056$ MS-SSIM, and $0.0095$ LPIPS, with a much smaller channel-adaptation module.","pith_inferences":["A natural extension not tested in the paper is to condition attention on variables other than SNR, such as bandwidth ratio, latency budget, or task type, since the CAQ formula only changes how queries are parameterized.","Real systems would quantize SNR feedback; the average-SNR variant suggests graceful degradation under feedback errors, but the paper only simulates Gaussian errors on the fading coefficient, leaving quantization effects untested.","Because the reported gains over SwinJSCC are moderate (around $0.1$-$0.6$ dB PSNR), their practical significance depends on whether they persist across seeds and channel realizations, which the single-run curves do not show.","The CAMHA block is a generic Transformer component, so the same mechanism could be inserted into Transformer coders for video, point clouds, or text, though the paper demonstrates only image transmission."],"forward_implications":["Transformer-based JSCC coders can be made channel-adaptive without a dedicated channel-adaptation network, by conditioning the attention queries and inputs on the channel state.","Because one model is trained over an SNR range, a single SNR-EQ-JSCC model can serve channel conditions from about $-10$ dB to $20$ dB without per-SNR retraining.","Under imperfect instantaneous SNR feedback, replacing instantaneous SNR with the average SNR at the per-block coders requires no retraining and still beats the baselines for moderate feedback errors.","The channel-adaptive query mechanism is credited specifically with improving perceptual metrics such as MS-SSIM and LPIPS, not just pixel-level PSNR.","In the reported setup, channel adaptation costs only $0.05\\%$ of SwinJSCC's Channel ModNet parameters and $6.38\\%$ of its FLOPs, so channel adaptation could be added to Transformer coders at negligible overhead."],"supporting_citations":[{"why":"Establishes the JSCC image-transmission problem that SNR-EQ-JSCC builds on and the CNN baseline lineage.","marker":"[1]"},{"why":"Supplies the Swin Transformer backbone and the Channel ModNet baseline whose storage and complexity are compared.","marker":"[2]"},{"why":"Provides the attention-module-based CNN baseline ADJSCC used in the comparisons and the imperfect-feedback experiment.","marker":"[8]"},{"why":"Defines the multi-head attention formulation whose queries, keys, and values are modified by CAQ and SNR embedding.","marker":"[11]"},{"why":"Provides the DIV2K dataset used for training and testing all compared schemes.","marker":"[12]"}],"fun_headline_variants":["SNR-embedded attention queries beat SwinJSCC with 0.05% overhead","Channel-adaptive queries cut adaptation cost to 0.05% and beat SwinJSCC","SNR-embedded attention improves PSNR and perception over SwinJSCC","Tiny SNR-adaptive attention module (0.05% storage) beats SwinJSCC","Channel-adaptive queries: 0.05% storage, 6.38% compute, better than SwinJSCC"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The conclusion that SNR-EQ-JSCC consistently outperforms SwinJSCC rests on single-run simulation comparisons with no reported variability, so the small measured gains could reflect training randomness rather than a systematic advantage.","fun_headline_variants_meta":{"raw":{"variants":["SNR-embedded attention queries beat SwinJSCC with 0.05% overhead","Channel-adaptive queries cut adaptation cost to 0.05% and beat SwinJSCC","SNR-embedded attention improves PSNR and perception over SwinJSCC","Tiny SNR-adaptive attention module (0.05% storage) beats SwinJSCC","Channel-adaptive queries: 0.05% storage, 6.38% compute, better than SwinJSCC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000756,"raw_usage":{"total_tokens":3428,"prompt_tokens":1078,"completion_tokens":2350,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":694,"completion_tokens_details":{"reasoning_tokens":2227}},"tokens_in":694,"tokens_out":2350,"duration_ms":16094,"temperature":1.0,"reasoning_tokens":2227,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:51:55.857975+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train SNR-EQ-JSCC and SwinJSCC on the same DIV2K split with several random seeds and compare the PSNR gap at $\\mu = 20$ dB and $r = 1/32$; if the gap over seeds is not consistently positive, or if the seed-to-seed spread exceeds the 0.32 dB gain, the claim of consistent outperformance is not established.","supporting_citations":[{"cited_title":"Deep joint source- channel coding for wireless image transmission,","cited_arxiv_id":null,"evidence_quote":"Establishes the JSCC image-transmission problem that SNR-EQ-JSCC builds on and the CNN baseline lineage."},{"cited_title":"SwinJSCC: Taming Swin Transformer for deep joint source-channel coding,","cited_arxiv_id":null,"evidence_quote":"Supplies the Swin Transformer backbone and the Channel ModNet baseline whose storage and complexity are compared."},{"cited_title":"Wireless image transmission using deep source channel coding with attention modules,","cited_arxiv_id":null,"evidence_quote":"Provides the attention-module-based CNN baseline ADJSCC used in the comparisons and the imperfect-feedback experiment."},{"cited_title":"Attention is all you need,","cited_arxiv_id":null,"evidence_quote":"Defines the multi-head attention formulation whose queries, keys, and values are modified by CAQ and SNR embedding."},{"cited_title":"Ntire 2017 challenge on single image super-resolution: Dataset and study,","cited_arxiv_id":null,"evidence_quote":"Provides the DIV2K dataset used for training and testing all compared schemes."}],"review_version":1}