{"id":"21a83538-33d2-44d2-868b-3a6728aa915c","arxiv_id":"2501.02738","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"SCSC inserts a learned preprocessor before standard BPG compression and a learned precoder/combiner around MIMO transmission, trained via a proxy network, to improve image transmission and downstream task performance in a standards-compatible way.","lead":"This paper proposes SCSC, a system that adds two small neural modules to standard image compression and MIMO transmission so that existing digital systems can gain some benefits of deep-learning joint source-channel coding. The authors report up to 29% bandwidth savings on image transmission with similar reconstruction quality and improved downstream vision task accuracy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proxy network fidelity and gradient alignment are asserted but never measured, leaving the claimed 29% bandwidth savings resting on an unvalidated surrogate.","rationale":"The reader's weakest assumption correctly identifies the proxy network as the load-bearing element of the proposed SCSC framework. The central claim — that PPEN and PCEN trained end-to-end via a proxy yield large bandwidth savings and task improvements over the true digital chain — depends entirely on the proxy being a faithful surrogate, both in its output distribution and in the gradients it provides. The paper explicitly recognizes this requirement in Sec. III.C ('If the discrepancy is too significant, the proxy network cannot substitute for the conventional codecs, indicating a failure in training') but never supplies the corresponding measurement. This is not a minor omission: the proxy is a continuous analog JSCC autoencoder, while the real system is a cascade of discrete, non-smooth blocks (BPG quantization, LDPC, QPSK, SVD precoding and MIMO detection). The gradients of these systems are fundamentally different, and the surrogate-gradient training used in Algorithm 2 has no correctness guarantee without an alignment check. The empirical results (Table II, Figs. 6–8) are plausible but cannot be independently reproduced from the paper, which lacks code, error bars, and the proxy-fidelity data. The paper's other contributions — the PPEN/PCEN architecture, the deep-unfolding formulation, and the extensive ablation/generalization studies — are solid engineering elements and receive due credit. However, the unvalidated proxy is the keystone: if it fails, the trained modules may not actually improve the real standards-based system, and the headline 29% bandwidth saving would be an artifact of optimizing against the wrong objective. The reader's CONDITIONAL verdict is therefore appropriate; it should not be changed to ACCEPT without the proposed fidelity check, nor REJECT because the framework is reasonable and the concern is empirically testable. The reader and this stress-test agree on the single most load-bearing concern, and the concrete test above would settle whether the proxy assumption holds.","tokens_in":21022,"tokens_out":6616,"duration_ms":64409,"concrete_test":"On a held-out set from Cityscapes/CVRG-Pano, for each (Q, LDPC rate, SNR) configuration used in the paper, compute (i) PSNR/MSE between proxy output \\hat{x}_a and true codec output \\hat{x}; (ii) the cosine similarity between the proxy gradient of the total loss w.r.t. a given PPEN input and a finite-difference estimate of the true gradient through the real codec (using the actual BPG/LDPC/QPSK/SVD chain as a black box) on a small batch. If the reconstruction discrepancy exceeds a few dB PSNR or the gradient cosine similarity falls below ~0.5, the proxy assumption is unsupported, and the trained modules should be re-evaluated on the real chain to confirm that the reported gains persist.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The entire SCSC training pipeline (Algorithm 2) replaces the non-differentiable digital chain with a JSCC autoencoder proxy (Sec. II.D, Fig. 5) and backpropagates the task loss through the proxy's output \\hat{x}_a while computing the actual loss on the real codec's output \\hat{x}. The paper states in Sec. III.C that 'if the discrepancy is too significant, the proxy network cannot substitute for the conventional codecs,' but it never quantifies this discrepancy for any (Q, LDPC rate, SNR) combination, nor does it verify that proxy gradients point in a valid descent direction for the true system. The proxy is a learned analog JSCC scheme; the real system is BPG (lossy block coding) + LDPC + QPSK + SVD/MIMO detection, whose distortions and Jacobians are very different (e.g., BPG's quantization yields zero gradients almost everywhere). Because PPEN and PCEN are optimized exclusively through this proxy, any bias in the proxy's input-output map or its gradients can create an SCSC that only performs well under the proxy, not on the actual standards-based codec. Without a proxy-fidelity check, the reported 29.46% BD-CBR savings (Table II) and mIoU gains (Fig. 8) could be training artifacts. The paper even acknowledges the necessity of a good match but provides no evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SCSC, a standards-compatible semantic communication framework for image transmission over MIMO channels. The system wraps conventional digital codecs (BPG/JPEG/JPEG2000 source coding, LDPC channel coding, QPSK modulation, SVD precoding) with two learnable modules: PPEN, a preprocessing network that filters the source image before compression, and PCEN, a deep-unfolded precoder/combiner enhancement network. Because the standard codecs are non-differentiable, training is performed through a proxy JSCC autoencoder that mimics the whole digital chain. The authors report PSNR, MS-SSIM, and semantic-segmentation mIoU results on Cityscapes, CVRG-Pano, and CLIC2021, claiming over 29% channel-bandwidth savings at equal PSNR (Table II) and gains in mIoU over several baselines, alongside robustness experiments across codecs, modulation orders, datasets, tasks, and imperfect CSI.","tokens_in":21354,"tokens_out":4230,"duration_ms":42858,"significance":"If the results hold, the framework is practically significant: it offers a path to semantic-communication gains while retaining deployment compatibility with existing digital standards, which is a genuine concern for end-to-end JSCC schemes. The PCEN design is a well-motivated ADMM-unfolded finite-alphabet precoder with a clean MSE formulation (Eq. (10)-(17)), and the quantization-adaptive PPEN architecture is a sensible way to handle multiple compression levels. The experimental breadth—multiple datasets, tasks, codecs, and channel models—is a strength. However, the central claim depends on an unvalidated proxy-network assumption, and the numerical evidence currently rests on single-run comparisons without error bars or code, which weakens confidence in the reported margins.","major_comments":[{"comment":"The entire end-to-end training pipeline replaces the non-differentiable digital chain with a JSCC autoencoder proxy and backpropagates through it, yet the paper never quantifies how faithfully the proxy represents the real BPG+LDPC+QPSK+SVD system. Section III.C acknowledges that 'if the discrepancy is too significant, the proxy network cannot substitute for the conventional codecs,' but no measurement of this discrepancy is provided for any (Q, LDPC rate, SNR) combination, and no gradient-alignment check is reported. Since PPEN and PCEN are optimized exclusively through proxy gradients (Algorithm 2), a biased proxy could yield modules that perform well only in simulation with the proxy and poorly when deployed on the actual standards-based chain. The authors should add a quantitative proxy-fidelity evaluation (e.g., output MSE/PSNR between proxy and real chain across the operating range) and ideally a gradient-alignment test or a comparison of SCSC fine-tuned on the real system vs. proxy-trained only.","section":"Section III.C and Algorithm 2"},{"comment":"All performance claims, including the headline 29.46% BD-CBR savings, are based on single-run experiments with no reported error bars, random seeds, or statistical significance. The margins over some baselines are modest (e.g., SCSC 56.37 vs. DeepJSCC-MIMO 54.24 mIoU in Table I; BD-PSNR 1.28 vs. 1.08 dB in Table II), so it is unclear whether the differences are meaningful. Furthermore, the training details for the baseline models (ADJSCC, DeepJSCC-MIMO, DSSCC) are incomplete—epochs, learning-rate schedules, SNR sampling, and segmentation-backbone training are not specified—making the comparisons difficult to reproduce. Please provide multi-seed results with variance or significance tests, and release code or detailed training configurations.","section":"Section IV.A.4 and Table II"},{"comment":"The text states that SCSC 'slightly underperforms compared to the DeepJSCC-MIMO method' in the mIoU comparison of Fig. 8, yet Table I reports SCSC mIoU 56.37 vs. DeepJSCC-MIMO 54.24 at R=1/6. These statements conflict. Please clarify the operating point used for Table I and reconcile it with the curves in Fig. 8; if Table I is at a different CBR or SNR, that must be stated explicitly.","section":"Section IV.B.3 and Table I"},{"comment":"The proxy network is pre-trained to mimic the standard codec before end-to-end training, but PPEN changes the input distribution of the proxy during SCSC training. The paper does not state whether the proxy is re-finetuned or updated within the K end-to-end epochs. If the proxy remains frozen while PPEN evolves, its fidelity may degrade, potentially making the training gradients increasingly stale. Please specify the proxy update schedule and report the proxy loss (Eq. (18)) during end-to-end training as a sanity check.","section":"Section II.D and Algorithm 2"}],"minor_comments":[{"comment":"There is a typo: 'Gound Truth' should be 'Ground Truth'.","section":"Figure 10 caption"},{"comment":"The notation is inconsistent: the equation defines ̄r_{i,c} and ^r_{j,c}, while the text refers to '¯ri,c and ^rj,c'. Please unify the notation.","section":"Eq. (8) and surrounding text"},{"comment":"The sentence beginning 'While our approach slightly underperforms compared to the DeepJSCC-MIMO method...' is a grammatical fragment; it should be merged with the previous sentence or rephrased.","section":"Section IV.B.3"},{"comment":"References [15] and [48] appear to be the same work (arXiv preprint and IEEE IoT Journal version). Please cite only one, or clearly indicate that they are different versions of the same paper.","section":"References [15] and [48]"},{"comment":"The annotations like '(41,1/6)' in the figures are not explained in the captions. Please add a note that these denote (BPG quantization parameter Q, LDPC code rate) combinations. Additionally, in Table II, the sign convention of BD-CBR (negative meaning bandwidth savings) should be stated explicitly.","section":"Fig. 6 and Table II"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely and practically important problem, and the proposed framework is novel in its combination of learnable preprocessing and deep-unfolded precoding around standard codecs. My main concern is the unvalidated proxy-network assumption, which is the load-bearing component of the training scheme; the authors must demonstrate proxy fidelity and, ideally, show that proxy-trained modules transfer to the real non-differentiable system. The lack of code and single-run results is also a significant reproducibility issue, especially because the reported gains over strong baselines are moderate. I would encourage the editor to request the proxy-fidelity study and multi-seed experiments before considering acceptance. I do not see this as a circularity problem, since the evaluation is against external baselines and standard codecs."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one if you want to see where the semantic-communication field is actually trying to deploy: a learned preprocessor (PPEN) in front of BPG, a deep-unfolding finite-alphabet precoder/combiner (PCEN) around MIMO, and a JSCC proxy network that supplies gradients through the non-differentiable digital chain. The integration is genuinely new; each block has ancestors, but the combination is not in the literature I know. The PCEN ADMM unfolding is clean, and the ablations support the qualitative role of each module.\n\nThe paper is also honest where it matters. Section III.C says the proxy must match the real codec, and Section II.D says the proxy replaces the codec during backprop. But that is exactly where the soft spot is. They never measure the proxy-codec discrepancy for any combination of quantization parameter, LDPC rate, and SNR, and they never verify that proxy gradients point in a useful descent direction for the real system. The proxy is a learned analog JSCC autoencoder; the real chain is BPG quantization plus LDPC/QPSK/SVD and MIMO detection. BPG's quantization has zero gradients almost everywhere, so the entire optimization of PPEN and PCEN rides on the proxy. If the proxy is biased, the 29% bandwidth saving and the mIoU gains could be artifacts of the surrogate rather than properties of the standards-based system. The paper acknowledges the requirement and then moves on without evidence. That is a load-bearing gap, not a cosmetic one.\n\nThe other weaknesses are the usual ones for this subfield: no code, no error bars, single-run curves, and some baseline descriptions are thin (ADJSCC is said to be suboptimal on high-res datasets, which softens the comparison). The throughput numbers in Table III are hard to interpret without knowing batch sizes and GPU details. None of these are fatal; they are fixable with a released implementation and more careful reporting.\n\nIf the proxy fidelity check comes back positive, this is a useful contribution for people working on 6G image transmission over legacy systems. If it comes back negative, the framework as trained is not doing what the curves claim. Either way, it deserves a serious referee. My recommendation: send it to review, ask for the proxy-fidelity experiment (quantify the distortion between the real codec output and proxy output across operating points, and ideally a gradient-alignment or fine-tuning sanity check), error bars on the headline numbers, and code release.","headline":"Solid standards-compatible semantic image transmission integration; the load-bearing question is the unmeasured fidelity of its proxy network.","tokens_in":21865,"tokens_out":2054,"would_cite":true,"duration_ms":20112,"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":"SCSC wraps the standard BPG/LDPC/QPSK/SVD digital chain in two learned modules, trains them through a proxy network, and reports over 29% channel-bandwidth savings at equal PSNR while improving semantic segmentation.","keywords":["semantic communication","joint source-channel coding","image transmission","MIMO","standard-compatible","deep learning","semantic segmentation","proxy network"],"falsifier":"Run the trained SCSC modules with the true BPG+LDPC+QPSK+SVD chain (no proxy) at the training SNRs and CBRs; if the reported bandwidth savings and mIoU gains vanish, or if the proxy's reconstructions at those operating points differ markedly from the real codec's, the proxy-training claim is falsified.","tokens_in":20842,"feed_emoji":"📡","tokens_out":9973,"duration_ms":89393,"temperature":0.7,"pith_summary":"SCSC is a semantic communication framework that keeps the conventional digital transmission chain—BPG compression, LDPC channel coding, QPSK modulation, SVD precoding—unchanged and adds two trainable modules around it: PPEN, which preprocesses the image before compression so that task-relevant semantics survive, and PCEN, which refines precoding and combining for finite-alphabet MIMO transmission. Because the digital chain is non-differentiable, the authors train the modules end-to-end through a proxy neural network that mimics the chain's input-output behavior during backpropagation. The paper reports that this standard-compatible system outperforms both separation-based digital baselines and an end-to-end learned JSCC baseline, saving more than 29% of channel bandwidth at the same PSNR and substantially improving semantic segmentation mIoU. The practical significance is that semantic communication could be deployed on existing systems by adding two modules rather than replacing codecs and standards.","feed_headline":"Two learned modules cut wireless image bandwidth by 29%","feed_subtitle":"Standard BPG and LDPC codecs stay untouched; small add-on networks beat end-to-end learned transmission.","key_machinery":"Three components carry the argument. PPEN (preprocessing-empowered network) is a distortion-aware image transform built from deformable convolution, strip pooling, and quantization-adaptive layers; it selects and reshapes information so the standard codec preserves what the downstream task needs. PCEN (precoder & combiner-enhanced network) is a deep-unfolded implementation of ADMM iterations for finite-alphabet precoding, with trainable step sizes and a learned linear combiner, which refines the transmitted and received signals under the constellation constraint. The proxy network is a JSCC autoencoder trained to imitate the full digital chain; it supplies gradients through otherwise discrete blocks while the true chain is used for forward propagation.","core_discovery":"On its own terms, the paper establishes that the advantage of learned JSCC over conventional digital transmission does not require abandoning the digital chain. It shows that a preprocessing network and a precoder/combiner-enhancement network, trained jointly through a differentiable proxy of the non-differentiable codec chain, can make the standard chain preserve task-relevant semantics and transmit them efficiently over MIMO channels. The central numerical claim is a 29.46% reduction in channel bandwidth at equal PSNR relative to the BPG+LDPC+QPSK baseline, a 1.28 dB PSNR gain at a fixed bandwidth ratio, and a semantic-segmentation mIoU that exceeds both the digital baseline and the learned JSCC baselines on panoramic images.","pith_inferences":["If the proxy-fidelity assumption holds, the same wrapper strategy could be applied to other non-differentiable transmission chains, such as polar-coded or OFDM systems, without redesigning the semantic modules.","PPEN can be read as a learned rate-distortion shaping operator that spends bits on task-relevant structures; connecting it to rate-distortion theory could produce predictable trade-offs between reconstruction fidelity and task accuracy.","Because the transmitted signal remains a finite-alphabet digital constellation, SCSC may keep the security and multi-hop robustness advantages of digital transmission over analog JSCC, a benefit the paper motivates but does not quantify.","A direct next experiment would be to measure the proxy-codec reconstruction gap and compare proxy gradients with finite-difference gradients of the true chain at every training SNR and CBR, giving a deployment rule for when the proxy must be retrained."],"forward_implications":["Legacy deployment requires adding only two modules: PPEN before the source encoder and PCEN around the MIMO precoder/combiner; the underlying codecs remain untouched.","At equal PSNR the system saves 29.46% channel bandwidth, and at a fixed bandwidth ratio of 1/12 it gains 1.28 dB in PSNR over the standard digital chain.","On the panoramic segmentation dataset the system raises mIoU from 30.62% (digital baseline) to 56.37%, and matches or exceeds learned JSCC baselines in most class-wise categories.","The gains transfer to JPEG and JPEG2000 source codecs, to the CLIC2021 dataset, to object detection and classification tasks, and to practical 5G channel models with imperfect CSI.","The deep-unfolded PCEN keeps the added complexity small enough that SCSC's throughput is about 19 times the digital baseline's in the reported setup."],"supporting_citations":[{"why":"Supplies the BPG source codec that PPEN wraps and that anchors the digital baseline.","marker":"[25]"},{"why":"Supplies the LDPC channel code used in the standard chain and baselines.","marker":"[28]"},{"why":"Provides the ADJSCC baseline whose bandwidth and PSNR results SCSC is compared against.","marker":"[47]"},{"why":"Provides the DeepJSCC-MIMO baseline, the end-to-end JSCC scheme SCSC approaches or surpasses at higher CBR.","marker":"[9]"},{"why":"Provides the DSSCC baseline used in the semantic segmentation mIoU comparisons.","marker":"[48]"},{"why":"Supplies the ADMM finite-alphabet precoding algorithm that PCEN unfolds into trainable layers.","marker":"[40]"},{"why":"Introduces deformable convolution, which PPEN's DAC module uses to handle object deformation.","marker":"[34]"},{"why":"Introduces strip pooling, which PPEN uses to capture global semantic context.","marker":"[35]"}],"fun_headline_variants":["Semantic add-ons cut wireless image bandwidth by 29%","Learned preprocessing and precoder cut 29% wireless image bandwidth","Standard codecs, learned add-ons: 29% less bandwidth for images","Add small networks to legacy codecs, save 29% on image bandwidth"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the proxy network is a faithful stand-in for the real digital chain, so that gradients computed through the proxy improve the true system; the paper does not measure the proxy-to-real discrepancy or check gradient alignment.","fun_headline_variants_meta":{"raw":{"variants":["Semantic add-ons cut wireless image bandwidth by 29%","Learned preprocessing and precoder cut 29% wireless image bandwidth","Standard codecs, learned add-ons: 29% less bandwidth for images","Add small networks to legacy codecs, save 29% on image bandwidth"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000742,"raw_usage":{"total_tokens":3288,"prompt_tokens":903,"completion_tokens":2385,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":519,"completion_tokens_details":{"reasoning_tokens":2305}},"tokens_in":519,"tokens_out":2385,"duration_ms":19430,"temperature":1.0,"reasoning_tokens":2305,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:05:59.574586+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained SCSC modules with the true BPG+LDPC+QPSK+SVD chain (no proxy) at the training SNRs and CBRs; if the reported bandwidth savings and mIoU gains vanish, or if the proxy's reconstructions at those operating points differ markedly from the real codec's, the proxy-training claim is falsified.","supporting_citations":[{"cited_title":"BPG image format,","cited_arxiv_id":null,"evidence_quote":"Supplies the BPG source codec that PPEN wraps and that anchors the digital baseline."},{"cited_title":"Design of low-density parity check codes for 5G new radio,","cited_arxiv_id":null,"evidence_quote":"Supplies the LDPC channel code used in the standard chain and baselines."},{"cited_title":"Wireless Image Transmission Using Deep Source Channel Coding With Attention Modules","cited_arxiv_id":"2012.00533","evidence_quote":"Provides the ADJSCC baseline whose bandwidth and PSNR results SCSC is compared against."},{"cited_title":"Deep joint source-channel coding for adaptive image transmission over MIMO channels,","cited_arxiv_id":null,"evidence_quote":"Provides the DeepJSCC-MIMO baseline, the end-to-end JSCC scheme SCSC approaches or surpasses at higher CBR."},{"cited_title":"Joint task and data oriented semantic communications: A deep separate source-channel coding scheme,","cited_arxiv_id":null,"evidence_quote":"Provides the DSSCC baseline used in the semantic segmentation mIoU comparisons."},{"cited_title":"Finite-alphabet precoding for massive MU-MIMO with low-resolution DACs,","cited_arxiv_id":null,"evidence_quote":"Supplies the ADMM finite-alphabet precoding algorithm that PCEN unfolds into trainable layers."},{"cited_title":"Deformable convolutional networks,","cited_arxiv_id":null,"evidence_quote":"Introduces deformable convolution, which PPEN's DAC module uses to handle object deformation."},{"cited_title":"Strip pooling: Rethinking spatial pooling for scene parsing,","cited_arxiv_id":null,"evidence_quote":"Introduces strip pooling, which PPEN uses to capture global semantic context."}],"review_version":1}