{"id":"93f2f6dc-507a-4c6b-b933-8a3a95f021c5","arxiv_id":"2506.16859","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A closed-loop system that joins Reed-Solomon coding with residual network updates maintains 100% symbol accuracy in multimode fiber video transmission under disturbance.","lead":"This paper combines Reed-Solomon error-correcting codes with a self-updating residual network to transmit images and video through a disturbed 100-meter multimode fiber. The authors report 100% symbol accuracy and zero frame loss in a 16x16 pixel H.265 video stream under manual disturbance and no temperature control.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The RS decoder's correction radius is the linchpin of the closed-loop claim, but code parameters and raw error rates are undisclosed, so 100% symbol accuracy may be carried by the decoder rather than by learned reconstruction.","rationale":"The reader's weakest assumption correctly identifies the RS error-correction capacity as the pivot of the paper's central mechanism. My stress-test converges on the same point and sharpens it: the RS decoder is used both as a training-label validator and as part of the reported accuracy measurement, so without code parameters and raw error statistics, the 99.95% effective-sample rate and the 100% symbol accuracy cannot be independently checked. This is not an internal inconsistency or a fundamental flaw; RS coding theory is standard and the experimental setup is plausible. But the claimed advantage over MMDN and the closed-loop error-accumulation suppression both rest on the assumption that corrected labels equal transmitted codewords. The paper's omission of n, k, t and raw-vs-corrected accuracy leaves that assumption untested. The correct disposition is conditional: accept the contribution as a promising demonstration, but require disclosure of the RS code parameters, raw prediction error counts, and a comparison of raw versus RS-corrected accuracy before the 100% symbol-accuracy claim can be taken at face value. Since this matches the reader's stated weakest assumption, the verdict should remain unchanged.","tokens_in":7131,"tokens_out":7355,"duration_ms":81154,"concrete_test":"Request the experimental logs for the 16x16 video run: RS code parameters (n, k, t), the per-codeword raw byte-error counts before RS decoding, and the ground-truth transmitted codewords. Then verify two things: (1) every one of the 99.95% 'effective training samples' had RS decoder output exactly equal to the true transmitted codeword, not merely equal to some valid codeword; (2) compute the theoretical miscorrection probability for the stated t and confirm it is negligible given the observed raw error distribution. If raw error counts exceed t in any logged window, or if a nontrivial fraction of RS decoder outputs differ from the ground-truth codewords, then the closed-loop label quality assumption fails and the headline accuracy is attributable to code redundancy or miscorrection rather than to the learned reconstruction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"RRSDN's central mechanism is the RS decoder acting as a gatekeeper: it must correct raw network predictions before those predictions become pseudo-labels for self-supervised updates. The paper reports that 99.95% of effective training samples had zero byte errors after RS decoding, but it never states the RS code length n, dimension k, or correction capacity t, nor the raw byte-error distribution before decoding. For an MDS code over GF(2^8), if a raw prediction vector is within Hamming distance t of a codeword other than the transmitted codeword, the decoder will silently miscorrect: it outputs a valid but wrong codeword. In that regime, the pseudo-label is a codeword, so the loss can be low and the closed loop can reinforce a wrong mapping without any outward sign of failure. The 100% symbol-accuracy figure is measured after the same RS decoder, so it cannot independently certify that the transmitted information was recovered. Without n, k, t and raw error counts, the 99.95% effective-sample rate cannot be distinguished from (i) a healthy margin where raw errors rarely approach t, or (ii) a low-rate code whose large correction radius hides a network that has not actually learned the channel. The distinction is load-bearing because the paper's headline claim is not merely that RS coding helps, but that the jointly optimized system achieves reliable long-term transmission with zero frame loss and 100% symbol accuracy. Reported accuracy after decoding, with no raw accuracy or code rate, does not resolve which of these two regimes was observed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RRSDN, a self-supervised residual network combined with Reed-Solomon error correction for real-time image and video transmission through multimode fiber. Information is byte-encoded, modulated onto binary amplitude masks, transmitted through a 100-meter MMF, and reconstructed by a multi-subnetwork ResNet whose outputs are corrected by an RS decoder before being used as pseudo-labels in a closed-loop training update. Experiments with 8×8 and 16×16 pixel patterns and an H.265 video stream report 100% symbol accuracy, zero frame loss, and substantial improvements over MMDN and StaticNN baselines under manually applied disturbances and without temperature control.","tokens_in":7353,"tokens_out":3647,"duration_ms":38313,"significance":"If the reported results hold, the paper makes a practically relevant contribution by demonstrating that a standard algebraic error-correcting code can act as a reliable gatekeeper in a self-supervised online learning loop for a dynamic scattering channel. The closed-loop 'prediction–correction–optimization' idea is a useful integration of discrete coding with continuous network training, and the 100-meter MMF experiments with real-time video go beyond static laboratory demonstrations. However, the paper currently omits the RS code parameters and raw error statistics, and the performance claims rest on single experimental runs, so the central result is not yet independently verifiable; with the missing information supplied, the contribution would be substantially strengthened.","major_comments":[{"comment":"The RS code parameters are never specified: the block length n, message length k, and error-correction radius t are not given anywhere in the text, despite the claim that the decoder corrects errors within a 'predefined error correction capacity t'. The reported 99.95% effective training samples with zero byte errors after RS decoding cannot be interpreted without knowing t and the raw byte-error distribution at the decoder input. For the closed-loop claim, the authors should report n, k, t, the raw symbol error rate before decoding, and the rate of silent miscorrection (outputs that are valid codewords different from the transmitted codeword); without these, the 100% symbol accuracy may be substantially carried by the RS code rather than by the learned channel reconstruction.","section":"Results, Dynamic Reconstruction and Error Correction"},{"comment":"All accuracy results appear to come from single experimental runs; no repeated trials, error bars, or confidence intervals are reported. The headline claims of 100% symbol accuracy and zero frame loss in the video experiment are based on one 70-second transmission, and the Fig. 2 curves are presented without indication of run-to-run variability. The authors should state the number of independent trials, report the mean and spread of the accuracy metrics across trials, and specify the duration represented by each curve in Fig. 2.","section":"Results, Information Transfer Accuracy"},{"comment":"The baseline systems are not sufficiently specified to make the comparison reproducible. MMDN is cited as reference [25] but its architecture, training procedure, and hyperparameters are not described in the manuscript or the supplementary material, and StaticNN is defined only as a CNN whose parameters are frozen after pre-training. The quantitative claim of an improvement of over 35 percentage points over MMDN cannot be verified without a complete description of both baselines. Please provide full architectural and training details, or make the baseline implementations available with the same experimental protocol.","section":"Results, Information Transfer Accuracy"},{"comment":"The paper's claim that the closed-loop system 'fundamentally blocks erroneous predictions from entering the training data pipeline' presumes that the RS decoder never miscorrects. In self-supervised operation there is no oracle for the transmitted symbols, so if a raw prediction is within distance t of a valid but incorrect codeword, the decoder will output a wrong pseudo-label and the loop can reinforce an incorrect mapping without any visible failure. The authors should provide evidence that miscorrection did not occur, for example raw error histograms, the fraction of decoder corrections that were later verified against known pilot symbols or pre-training data, and a discussion of how close the operating point is to the decoder's correction boundary given the reported 99.95% effective-sample rate.","section":"Discussion and Closed-Loop Training"}],"minor_comments":[{"comment":"The text contains numerous typographical and formatting issues: 'Multimodefiber' appears without a space in the Introduction, the Fig. 2 caption includes the garbled phrase 'BECSSIM and PCC BAR', and symbols such as '?? 28' and '? = 0.4' should be 'GF(2^8)' and a dropout rate, respectively.","section":"Throughout"},{"comment":"The metric terminology is inconsistent: the abstract and results use 'symbol accuracy', while the discussion and figure captions refer to 'byte errors', 'BER', and 'BAR/BEC'. Please define all acronyms and use a single consistent set of accuracy metrics.","section":"Results and Discussion"},{"comment":"The sentence 'The pre-training phase lasted for the first 4 seconds and the first 20 seconds' is ambiguous: it likely means 4 seconds for the 8×8 case and 20 seconds for the 16×16 case, but this is not stated explicitly and should be clarified.","section":"Results, Information Transfer Accuracy"},{"comment":"The acronyms BAR and BEC are not defined in the caption or the main text; please explain what they stand for and how they are computed.","section":"Fig. 2 caption"}],"recommendation":"major_revision","confidential_remarks":"The central idea is timely and the experimental demonstration is potentially valuable, but the missing RS code parameters and raw error statistics, together with the absence of repeated trials, are load-bearing gaps that currently prevent assessment of the core reliability claim. These are fixable within the scope of the manuscript, and I would encourage a revision that supplies the requested details rather than a rejection. If the authors can show that the operating point is far from the RS miscorrection boundary and provide trial-to-trial reproducibility, the paper could become acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The interesting thing here is the loop: a Reed-Solomon decoder sits between the network's raw predictions and the pseudo-labels used for self-supervised updates, and gradients are passed back through the decoder. I haven't seen that exact combination before. The demonstration on a 100 m MMF with manually induced lateral disturbances is real, and the effective training sample rate going from 10.09% to 99.95% is a concrete sign that the RS gatekeeper is doing useful work. Giving it credit where due: the architecture is sensible, the video transmission test with zero frame loss is a decent proof-of-concept, and the comparison to the prior MMDN is at least directionally informative.\n\nThe problems are all about transparency, not about the core idea being wrong. The paper never states the RS code parameters—n, k, t—nor the raw byte-error distribution before decoding. That is load-bearing. The headline 100% symbol accuracy and the 99.95% effective-sample rate are both measured after RS decoding. Without the code rate and the raw error rate, you cannot tell whether the network learned the channel or whether a low-rate code is simply correcting almost everything. The closed-loop training depends on the raw prediction errors staying inside the decoder's correction radius; if they don't, the decoder silently miscorrects and feeds wrong codewords back as labels. The paper doesn't show how close it is to that boundary. The gradient passthrough through the RS decoder is also asserted but not described—RS decoding is not differentiable, so a straight-through estimator or similar needs to be stated. Minor but real: no error bars or repeated trials, the 16×16 video is only 256 pixels, and the data and code are unavailable.\n\nThese are fixable gaps. The math is standard coding theory, and the experimental claims are internally consistent. I don't think the paper is hiding a fatal flaw; it is just under-reporting the quantities that would let a reader know how much of the gain is due to error correction versus learned reconstruction. The authors should be asked to supply the RS code parameters, the raw symbol accuracy before RS decoding, and a few independent runs.\n\nFor you: this is a good reading-group paper for the discussion of self-supervised loops in physical channels, and a serious editor should send it to peer review rather than desk reject. With the missing parameters disclosed, it could be a solid contribution to the MMF imaging/communication subfield.","headline":"A genuinely new closed-loop idea—RS-corrected pseudo-labels feeding a self-supervised update—with a plausible 100 m MMF demo, but the missing RS parameters and raw error rates leave the headline 100% accuracy uninterpretable as learning versus coding.","tokens_in":7961,"tokens_out":1751,"would_cite":false,"duration_ms":20219,"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":"Pairing Reed-Solomon error correction with a self-updating residual network lets a 100-meter multimode fiber carry a 16x16-pixel H.265 video stream with zero frame loss and 100% symbol accuracy despite manual disturbances.","keywords":["multimode fiber","Reed-Solomon code","residual network","self-supervised learning","error accumulation","speckle reconstruction","H.265 video transmission"],"falsifier":"Count corrected-label failures during a longer or harsher run: record the number of byte errors per RS block before decoding and compare with the code's correction capacity $t$ (which the paper should report). If the raw error count exceeds $t$ in any block, the RS decoder will mis-correct and feed a wrong label into the self-supervised loop; the claimed suppression of error accumulation then predicts a measurable accuracy drop after that block. A reader could also check the claim that 99.95% of effective training samples had zero byte errors after RS decoding by re-deriving this fraction from the raw error distribution.","tokens_in":6858,"feed_emoji":"📡","tokens_out":9980,"duration_ms":92495,"temperature":0.7,"pith_summary":"This paper tries to show that the two classic failure modes of information transmission through multimode optical fiber—speckle distortion from modal dispersion and error accumulation when a self-learning receiver drifts—can be handled together by treating error-correcting coding as part of the neural network's training loop. It proposes RRSDN, a residual convolutional network whose output labels pass through a Reed-Solomon decoder before they are used as training targets, so wrong predictions are corrected before they can poison the next update. The experimental claim is that on a 100-meter multimode fiber with manual lateral disturbances and no temperature control, this closed loop keeps symbol accuracy above 99.99 percent and transmits a 16x16-pixel H.265 video stream with no frame loss, while a comparison dynamic network's accuracy drops to about 65 percent. If true, it matters because it offers a path to reliable long-term optical links in environments that change faster than traditional transmission-matrix calibration can follow.","feed_headline":"Video survives disturbed multimode fiber with RS-coded neural loop","feed_subtitle":"Reported result: 99.99% symbol accuracy and complete H.265 video recovery without frame loss.","key_machinery":"The load-bearing mechanism is the closed-loop training pipeline of RRSDN, with the RS decoder placed between network prediction and pseudo-label creation. RS codes operate on 8-bit symbols in the Galois field $\\mathrm{GF}(2^8)$: the encoder adds parity bytes at the transmitter, and the decoder locates and corrects byte errors up to its correction capacity $t$ before a prediction becomes a training label. That block prevents wrong pseudo-labels from entering the self-supervised update, while subnetworks S1 and S2 (rebuilt every five time windows) handle transient disturbances and the frozen long-term subnetwork L models slow drift; an uncertainty-weighted fusion combines their outputs. The resulting cycle of prediction, correction, and optimization is what lets discrete coding cooperate with continuous gradient training.","core_discovery":"The central claim is that integrating Reed-Solomon (RS) error-correcting coding into a residual-network-based self-supervised reconstruction pipeline suppresses the error accumulation that otherwise destabilizes adaptive multimode-fiber transmission. The transmitter RS-encodes each information byte sequence and maps it to binary amplitude masks displayed on a digital micromirror device; the receiver reconstructs the mask from speckle, applies an RS decoder over $\\mathrm{GF}(2^8)$ to correct byte errors, and feeds the corrected symbols back through a gradient passthrough to update the network. Over 300-second runs on a 100-meter multimode fiber with $\\pm 0.1$ mm lateral perturbations and no temperature control, RRSDN is reported to keep average symbol accuracy above $99.99\\%$ for both 8×8 and 16×16 patterns, and to beat the MMDN baseline by 35.01 percentage points in the 16×16 case (MMDN: $64.94\\%$). In the 70-second video test it recovered the complete H.265 stream from 28,000 mask frames with zero frame loss, while MMDN lost frames and StaticNN failed after 18 seconds.","pith_inferences":["A natural extension is rate-adaptive coding: the paper's own repair mechanism allows the transmitter to monitor channel stability, for example via SSIM fluctuation, and vary RS redundancy to trade bandwidth for reliability; this is suggested in its discussion but not demonstrated.","The same closed-loop correction idea may transfer to other scattering channels, such as tissue or underwater media, where self-supervised receivers face the same pseudo-label poisoning, but the paper only tests multimode fiber.","If the result holds, the practical ceiling of the method is set by the RS correction capacity: scaling to kilometer-scale fibers with nonlinear effects would require either much stronger codes or an outer error-correction layer, since raw error bursts will eventually exceed any fixed $t$."],"forward_implications":["A 100-meter multimode fiber under manual disturbance can carry an H.265-compressed 16×16 video stream with zero dropped frames and 100% symbol accuracy, with reconstructed-frame SSIM close to 1.0.","The fraction of effective training samples with zero byte errors rises from 10.09% for the MMDN baseline to 99.95% for RRSDN, and the method converges in about 100 iterations, 66% faster than MMDN.","Error accumulation in self-supervised MMF reconstruction can be blocked at the label source, rather than only mitigated by architecture design or multi-scenario joint training.","Because the RS encoder and decoder are separate from the network, the approach is compatible with existing spatial light modulation hardware and camera-based receivers."],"supporting_citations":[{"why":"Provides the MMDN baseline; its self-supervised reconstruction is the comparison RRSDN must beat and whose error-accumulation failure motivates the closed-loop design.","marker":"[25]"},{"why":"Describes the dynamic memory/self-supervised framework for unstabilized diffusive media, the immediate predecessor that the paper extends with RS correction.","marker":"[26]"},{"why":"Supplies the residual learning blocks that make up each subnetwork in the reconstruction stage.","marker":"[27]"},{"why":"Supplies the Reed-Solomon decoder architecture with burst-error correction, the mechanism that corrects predictions before they become training labels.","marker":"[28]"},{"why":"Showed that deep learning can reconstruct input images from speckle through a dynamically perturbed multimode fiber, establishing the task RRSDN performs.","marker":"[20]"},{"why":"Provides the 100-meter MMF multi-disturbance joint-training framework whose static generalization limits motivate real-time adaptation.","marker":"[22]"},{"why":"Established transmission-matrix predictability of MMF; its environmental sensitivity is the practical limitation RRSDN addresses.","marker":"[6]"}],"fun_headline_variants":["RS-coded neural loop stabilizes video through disturbed fiber","Residual Reed-Solomon network keeps multimode fiber video intact","Zero-loss H.265 video over 100m fiber with RS coding","Deep learning plus Reed-Solomon codes beat fiber disturbances"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The raw prediction errors made by the network must always stay within the Reed-Solomon decoder's correction capacity $t$; if a block contains more byte errors than $t$, the decoder mis-corrects and injects a wrong label into the self-supervised training loop, and the paper does not state the code rate or $t$.","fun_headline_variants_meta":{"raw":{"variants":["RS-coded neural loop stabilizes video through disturbed fiber","Residual Reed-Solomon network keeps multimode fiber video intact","Zero-loss H.265 video over 100m fiber with RS coding","Deep learning plus Reed-Solomon codes beat fiber disturbances"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000237,"raw_usage":{"total_tokens":1549,"prompt_tokens":1028,"completion_tokens":521,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":644,"completion_tokens_details":{"reasoning_tokens":449}},"tokens_in":644,"tokens_out":521,"duration_ms":5426,"temperature":1.0,"reasoning_tokens":449,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:16:31.108808+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Count corrected-label failures during a longer or harsher run: record the number of byte errors per RS block before decoding and compare with the code's correction capacity $t$ (which the paper should report). If the raw error count exceeds $t$ in any block, the RS decoder will mis-correct and feed a wrong label into the self-supervised loop; the claimed suppression of error accumulation then predicts a measurable accuracy drop after that block. A reader could also check the claim that 99.95% of effective training samples had zero byte errors after RS decoding by re-deriving this fraction from the raw error distribution.","supporting_citations":[],"review_version":2}