{"id":"0e727c28-7923-4a2f-8e10-75bd7574b95e","arxiv_id":"2509.01119","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A Barlow-Twins-style self-supervised encoder is applied to goal-oriented semantic communication, claiming robust classification over noisy channels, though the headline gain is contradicted by its own aggregate table.","lead":"SC-GIR trains an image encoder with a covariance-based self-supervised loss, then sends compressed latent codes over noisy wireless channels for classification. It reports strong accuracy at high compression, but its headline 'nearly 10%' gain does not hold in its own averaged table.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported gains may be an artifact of unequal channel-symbol and power budgets: the paper never specifies the channel codec, k/n mapping, or power normalization for SC-GIR versus baselines.","rationale":"The reader's weakest assumption is the same as my primary concern. I agree because even the internal Table V inconsistency, while important, could be a typo; the fair-comparison gap is a structural omission that would invalidate the central empirical claim regardless of table typos. The paper's description of the channel chain is too vague to verify that all schemes operate at equal bandwidth and power. This is the most load-bearing condition for the headline claim. The proposed check—recomputing actual symbol counts and powers—would settle it. If the budgets are matched and gains persist, the claim is credible; if not, the claim fails. Therefore the reader's CONDITIONAL verdict is appropriate; no change.","tokens_in":23019,"tokens_out":10366,"duration_ms":115452,"concrete_test":"Obtain from the authors (or from released code) the exact channel encoder/decoder architecture and power normalization; then recompute the actual channel-symbol count and average symbol power per image for each method at k/n=0.1 on CIFAR-10. If SC-GIR transmits more than 307 symbols per image (the stated size for k/n=0.1) or uses higher average power than DeepJSCC/SemCC, the Fig. 5 comparison is not resource-equivalent and the claimed gain is an artifact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that SC-GIR outperforms baselines by nearly 10% and exceeds 85% accuracy under compression/noise depends on a controlled resource comparison. The paper never specifies: (i) the channel encoder/decoder architectures (Section III-D is qualitative; no trained/fixed codec is described), (ii) how the reported k/n ratios (e.g., 0.1–0.6, transmitted sizes 307–1,843 bytes for CIFAR-10) are realized for a semantic encoder that outputs a 2048-dimensional latent, and (iii) the transmit power normalization used to set SNR in Eq. (1). If SC-GIR transmits more channel symbols or at higher average power at a given nominal k/n, the accuracy advantage in Figs. 5–6 reflects unequal resource budgets rather than better invariant representations. Supporting evidence that the numbers need scrutiny: in Table V, the 'Average' column is inconsistent with the per-dataset entries (SC-GIR's row average is ~88.3, not 66.6, and the reported 66.6 is not ~10% above DeepJSCC/DeepSC's 66.7).","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SC-GIR, a self-supervised semantic communication framework for image transmission. A ResNet-34 encoder with a multi-layer projection head is trained with a Barlow-Twins-style cross-correlation loss on two augmented views, producing a compressed latent that is transmitted over AWGN/Rayleigh channels and used for downstream classification. Experiments on CIFAR-10/100, MNIST/FMNIST, STL-10, Flower-17, Cityscapes, and PACS are reported, and the abstract claims nearly 10% improvement over baselines and over 85% classification accuracy at low SNR/compression. The paper also includes segmentation and domain-generalization experiments as evidence of task-agnostic representation quality.","tokens_in":23396,"tokens_out":3683,"duration_ms":41994,"significance":"If the resource-controlled comparison and the reported numbers were correct, the paper would be a useful demonstration that a self-supervised covariance-based encoder can serve as a task-agnostic semantic source for wireless image classification, reducing reliance on labeled data and joint end-to-end training. The claim of robust performance at low compression ratios and low SNR is practically significant for IoT/edge communications. However, the manuscript currently contains several load-bearing inconsistencies: the headline average accuracy in Table V is not the mean of the reported rows, Eq. (3) is not a valid information-theoretic identity, and the channel-symbol/power budgets against baselines are not specified. These issues must be resolved before the central claims can be accepted.","major_comments":[{"comment":"The abstract claims SC-GIR outperforms baselines by 'nearly 10%', but Table V's reported rows contradict this. The SC-GIR row average is (87.2+98.0+99.3+85.5+86.5+73.1)/6 ≈ 88.3, not 66.6; DeepJSCC's row average is ≈ 73.7, not 66.7. The 'Average' column is therefore not the mean of the per-dataset entries. The Section V-B discussion ('average accuracy of 66.6% ... nearly matching DeepJSCC and DeepSC, both at 66.7%') would imply no gain, not a 10% gain. Please recompute the table, restate the headline claim, and reconcile the abstract with the corrected numbers.","section":"Abstract and Table V"},{"comment":"The decomposition I(X;S) = I(X;S|Y) + I(S;Y) is not an identity in general. The chain rule gives I(X;S) = I(X;S|Y) + I(S;Y) only under additional assumptions (e.g., some form of conditional independence/dependence structure that is not stated). This equation is the formal justification for the redundant/task-related split and for the subsequent causal model. Please replace it with a correct decomposition or explicitly state and justify the assumptions under which it holds; otherwise the theoretical motivation for the loss is unsupported.","section":"Eq. (3), Section III-B"},{"comment":"The comparison with baselines is not resource-controlled as reported. The paper never specifies the channel encoder/decoder architecture (Section III-D is qualitative), how the 2048-dimensional semantic latent is mapped to the stated k/n ratios and transmitted byte sizes (307–1,843 bytes), or the transmit power normalization used to set SNR in Eq. (1). Different methods may therefore transmit different numbers of channel symbols or at different average power for the same nominal k/n. This directly affects the validity of Figs. 5–6 and the 'outperforms baselines' claims. Please specify the full transmission chain, including codec, symbol mapping, and power normalization, and verify that all methods use equal channel resources.","section":"Section III-D and Section V-C"},{"comment":"The paper states that the IB objective in Eq. (13) is 'reformulated' into the cross-correlation loss in Eq. (10) 'through simplifications and approximations [40]'. This is not demonstrated and is questionable: Eq. (10) is the Barlow Twins loss, a heuristic covariance-based objective, and no derivation connecting the IB Lagrangian to the diagonal/off-diagonal cross-correlation terms is given. Please either provide a concrete derivation or reframe Eq. (10) as an empirically motivated loss, without claiming it is derived from IB.","section":"Section IV-B, Eqs. (10) and (13)"},{"comment":"The generalization claims in Section V-F are not supported by sufficient experimental detail. For the Cityscapes segmentation experiment, the paper does not state the decoder architecture, training protocol, resolution, or how SC-GIR's semantic encoder is integrated; for PACS, the fine-tuning/evaluation protocol is not given. Without these details, the strong mIoU and domain-generalization numbers cannot be assessed. Please add the missing protocol information or temper the generalization claims accordingly.","section":"Section V-F, Tables VIII-IX"}],"minor_comments":[{"comment":"The algorithm does not match Eq. (10): the λ weighting on the off-diagonal term is missing, and the definitions of Lon and Loff (e.g., line 8) are unclear—the notation 'C − fdiag(C) + 1' is not a standard way to target off-diagonal entries. Please align the pseudocode with the equation.","section":"Algorithm 1"},{"comment":"The text says 'The first component of the cross-correlation focuses on the off-diagonal loss', but Eq. (10)'s first term is the diagonal term (1−Cii)². This appears to be a wording error.","section":"Section IV-B, paragraph after Eq. (10)"},{"comment":"The Rayleigh subplot caption says 'BPG 12 rate LDPC' while the text refers to 'BPG 3/4 rate LDPC'. Please correct the inconsistency.","section":"Fig. 5 caption"},{"comment":"The metric description says 'High cosine similarity between the original data and encoded representations', but Fig. 8 shows cosine similarity between latent representations of the two augmented views, not between original data and encoding. Please clarify.","section":"Section V-A, Metrics"},{"comment":"Reference [59] appears to duplicate reference [26]; both cite the same 'Contrastive learning-based semantic communications' paper. Please consolidate.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The central conceptual element—the cross-correlation loss—is effectively Barlow Twins applied to a semantic communication setting, so the novelty rests mainly on the application and the experimental comparison. The current manuscript's own Table V undermines the headline performance gain, and the missing resource-budget specification is a serious correctness risk for the wireless experiments. These are fixable with a corrected table, a clear transmission-chain description, and a rerun if needed, but they should be addressed before sending to review again. I would also appreciate editorial attention to the unusual mismatch between the abstract/conclusion (which emphasize CIFAR-10, STL-10, Flower-17) and the actual experiments, which include many more datasets; this is cosmetic but contributes to the impression of rushed preparation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version. The paper is an incremental but sensible application of Barlow Twins to goal-oriented semantic communication, with a broad benchmark. The central claim—nearly 10% average improvement and over 85% accuracy—does not hold up against the paper's own Table V, and the comparison protocol is too underspecified to validate the gains.\n\nWhat's actually new: they train a task-agnostic encoder with a covariance-based cross-correlation loss (Eq. 10), which is Barlow Twins, and plug it into a standard JSCC-style pipeline. That's a legitimate and potentially useful move: it removes the need for joint transceiver training and labeled data, and it keeps inference cheap. The experimental scope is genuinely large—eight datasets, SNR sweeps, compression ratios, a domain generalization study, and a segmentation transfer test. The CIFAR-10 SNR curves show a consistent advantage, which is a meaningful data point if the resource budgets are comparable.\n\nThe soft spots are real and load-bearing. First, Table V is internally inconsistent: the average column for SC-GIR is 66.6, but the row mean is ~88.3, and the text says SC-GIR nearly matches DeepJSCC/DeepSC at 66.7. That contradicts the abstract's \"nearly 10%\" and \"over 85%\" claims. Second, Eq. (3) is not a valid decomposition of mutual information; the identity I(X;S)=I(X;S|Y)+I(S;Y) is generally false, which undercuts the information bottleneck story. Third, and most important, the channel codec is never specified: no architecture, no training procedure, no description of how k/n is computed for a 2048-dimensional latent versus DeepJSCC, SemCC, or BPG+LDPC, and no power normalization. Without that, the gains in Figs. 5–6 could be an artifact of unequal symbol or power budgets. Fourth, the segmentation and PACS results lack protocol details (decoder head, training data, hyperparameters), making them hard to assess.\n\nThe paper is for researchers looking for a simple self-supervised encoder baseline for task-oriented SC. It's not a reorganizing result, but the idea and the benchmark are worth having. It deserves peer review, but only with major revision: fix Table V and Eq. (3), specify the codec and power normalization precisely, and release code. If those are addressed, it could become a useful reference.","headline":"A sensible but incremental application of Barlow Twins to semantic communication; the benchmark is broad and the SNR curves are suggestive, but the paper's own numbers are inconsistent and the channel codec protocol is too underspecified to support the headline claims.","tokens_in":23802,"tokens_out":5147,"would_cite":false,"duration_ms":56234,"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":"SC-GIR claims a self-supervised, covariance-based encoder can extract task-essential image features, transmit them over noisy channels, and still classify above 85% accuracy at high compression.","keywords":["semantic communication","goal-oriented communication","invariant representation learning","self-supervised learning","cross-correlation loss","joint source-channel coding","wireless image transmission","redundancy reduction"],"falsifier":"At k/n = 0.1 and SNR = 5 dB, measure the actual transmitted energy per image and the number of channel uses for SC-GIR, DeepJSCC, SemCC, SemRE, and BPG+LDPC, then rerun the classification comparison with exactly matched power and bandwidth. If SC-GIR's 85% AWGN and 80% Rayleigh accuracy advantage disappears under matched resource budgets, the headline outperformance claim is an artifact of unequal comparison rather than a property of the learned representation.","tokens_in":22987,"feed_emoji":"📡","tokens_out":4814,"duration_ms":59069,"temperature":0.7,"pith_summary":"The paper tries to show that a wireless transmitter can send only the task-relevant part of an image, learned without any labels, and still let a receiver classify it accurately. It introduces SC-GIR, a semantic communication framework whose encoder is trained with a cross-correlation loss on two augmented views of each image, forcing the latent representation to be invariant to distortions while discarding redundant features. If this works, IoT and edge devices could communicate meaning rather than raw pixels, saving bandwidth and power without needing per-task labeled training data. The paper reports that SC-GIR outperforms baseline semantic and traditional codecs by nearly 10% on average and maintains over 85% classification accuracy on CIFAR-10 under AWGN at a 0.1 compression ratio.","feed_headline":"Semantic codec keeps 85% accuracy on noisy links","feed_subtitle":"SC-GIR sends only task-relevant image features, learned without labels, beating DeepJSCC and SemCC at the same compression.","key_machinery":"The load-bearing mechanism is the cross-correlation loss on standardized embeddings: L_cross-corr = sum_i (1 - C_ii)^2 + λ sum_{i≠j} (C_ij)^2, where C is the empirical cross-correlation matrix between the outputs of two parallel encoders fed with two augmented views of the same image. The diagonal terms push the representation to be invariant to the augmentations, and the off-diagonal terms push different feature dimensions to be decorrelated, which the paper interprets as redundancy reduction aligned with the information bottleneck principle. This loss is computed during training only; at inference, a single non-augmented image is encoded once and transmitted, so the framework adds no infer","core_discovery":"The central claim is that a goal-oriented semantic communication system can be built from an invariant representation learned purely by self-supervision: two distorted views of the same image are fed through the same encoder, and a covariance-based cross-correlation loss makes the embedding stable across views while decorrelating its dimensions. This yields a compressed latent that survives Rayleigh fading and AWGN well enough for downstream classification, without requiring joint training of transmitter and receiver or labeled data. The paper further claims that the same representation transfers to semantic segmentation and domain generalization, reporting 63.5 mean IoU on Cityscapes and 76","pith_inferences":["If the representation is truly task-agnostic, one could test it by attaching multiple downstream heads (classifier, segmenter, detector) to the same transmitted latent and measuring whether all benefit without per-task retraining of the encoder; the paper evaluates classification and segmentation separately but not simultaneously.","The cross-correlation loss depends on batch statistics for standardization, so a natural stress test is to evaluate SC-GIR with very small or non-i.i.d. batches, which are common in real-time edge inference; the paper's experiments use larger batches and do not address this failure mode.","The 'invariant' claim could be probed by causally shifting spurious correlations in the source data, e.g., changing background or lighting while keeping the label fixed; PACS is a useful domain-shift proxy but not a causal intervention, so the invariance claim remains partially open.","The paper suggests the framework extends to text and audio; since the covariance loss is modality-agnostic, a concrete next step would be a multimodal benchmark with the same Rayleigh channel model, testing whether the same redundancy-reduction principle transfers across signal types."],"forward_implications":["If SC-GIR is correct, semantic communication no longer requires labeled training data or joint transmitter-receiver training, removing a major barrier to deployment in dynamic IoT environments.","The reported accuracy at k/n = 0.1 suggests that a tenfold bandwidth reduction is possible for classification-oriented image transmission while keeping task accuracy above 80-85% on common benchmarks.","Because the same representation supports classification, segmentation, and domain-generalization tasks, the framework points toward one encoder serving many downstream goals at the receiver rather than a dedicated codec per task.","The training-time-only augmentation scheme means the learned invariance comes at no extra inference cost, making the approach compatible with resource-constrained edge devices.","If the domain-generalization results hold, SC-GIR could reduce or eliminate the need to retrain the semantic encoder when a deployment environment's visual style changes."],"supporting_citations":[{"why":"Supplies the DeepJSCC baseline and the joint source-channel coding paradigm that SC-GIR compares against and seeks to improve.","marker":"[10]"},{"why":"Supplies the SemCC and SemRE contrastive-learning baselines whose performance SC-GIR is measured against.","marker":"[59]"},{"why":"Supplies the covariance-based redundancy-reduction principle that underlies the SC-GIR cross-correlation loss.","marker":"[39]"},{"why":"Supplies the exact form of the cross-correlation loss and the simplification connecting the information bottleneck objective to this loss.","marker":"[40]"},{"why":"Provides the theoretical criterion that a representation S is minimally sufficient for task Y when I(X;Y|S)=0, used to justify the objective.","marker":"[32]"},{"why":"Provides the decomposition of I(X;S) into task-related and redundant information, framing the redundancy-reduction goal.","marker":"[33]"},{"why":"Supplies the information bottleneck principle that motivates the trade-off between compression and task-relevant information.","marker":"[35]"},{"why":"Supplies the structured causal model used to separate causal task-related features from non-causal spurious features.","marker":"[36]"}],"fun_headline_variants":["Self-supervised codec keeps 85% accuracy on noisy links","No-label sc codec beats baselines by 10% on noisy channels","Invariant codec learns without labels or joint training","Semantic compression hits 85% accuracy on noisy channels","Self-supervised codec transfers to segmentation and classification"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The reported accuracy gains assume that every method transmits the same number of channel symbols under the same power and bandwidth budget at each compression ratio, but the paper never specifies the channel encoder/decoder architectures or power normalization used for the baselines.","fun_headline_variants_meta":{"raw":{"variants":["Self-supervised codec keeps 85% accuracy on noisy links","No-label sc codec beats baselines by 10% on noisy channels","Invariant codec learns without labels or joint training","Semantic compression hits 85% accuracy on noisy channels","Self-supervised codec transfers to segmentation and classification"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000983,"raw_usage":{"total_tokens":4000,"prompt_tokens":728,"completion_tokens":3272,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":472,"completion_tokens_details":{"reasoning_tokens":3197}},"tokens_in":472,"tokens_out":3272,"duration_ms":28893,"temperature":1.0,"reasoning_tokens":3197,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T12:52:19.631332+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"At k/n = 0.1 and SNR = 5 dB, measure the actual transmitted energy per image and the number of channel uses for SC-GIR, DeepJSCC, SemCC, SemRE, and BPG+LDPC, then rerun the classification comparison with exactly matched power and bandwidth. If SC-GIR's 85% AWGN and 80% Rayleigh accuracy advantage disappears under matched resource budgets, the headline outperformance claim is an artifact of unequal comparison rather than a property of the learned representation.","supporting_citations":[{"cited_title":"Deep joint source- channel coding for wireless image transmission,","cited_arxiv_id":null,"evidence_quote":"Supplies the DeepJSCC baseline and the joint source-channel coding paradigm that SC-GIR compares against and seeks to improve."},{"cited_title":"Contrastive learning-based semantic communications,","cited_arxiv_id":null,"evidence_quote":"Supplies the SemCC and SemRE contrastive-learning baselines whose performance SC-GIR is measured against."},{"cited_title":"VICReg: Variance-invariance- covariance regularization for self-supervised learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the covariance-based redundancy-reduction principle that underlies the SC-GIR cross-correlation loss."},{"cited_title":"Barlow twins: Self-supervised learning via redundancy reduction,","cited_arxiv_id":null,"evidence_quote":"Supplies the exact form of the cross-correlation loss and the simplification connecting the information bottleneck objective to this loss."},{"cited_title":"Learning robust representations via multi-view information bottleneck,","cited_arxiv_id":null,"evidence_quote":"Provides the theoretical criterion that a representation S is minimally sufficient for task Y when I(X;Y|S)=0, used to justify the objective."},{"cited_title":"To compress or not to compress—self- supervised learning and information theory: A review,","cited_arxiv_id":null,"evidence_quote":"Provides the decomposition of I(X;S) into task-related and redundant information, framing the redundancy-reduction goal."},{"cited_title":"Task-oriented communication with out-of-distribution detection: An information bottleneck framework,","cited_arxiv_id":null,"evidence_quote":"Supplies the information bottleneck principle that motivates the trade-off between compression and task-relevant information."},{"cited_title":"Discovering Invariant Rationales for Graph Neural Networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the structured causal model used to separate causal task-related features from non-causal spurious features."}],"review_version":1}