{"id":"74141054-bec4-4758-be18-3dbc8f43c00a","arxiv_id":"1909.05147","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A simulated DNN blind detector approaches perfect-CSI ML error rates in FSO-MIMO, but the multiuser model and missing artifacts undercut the claim.","lead":"This paper trains deep neural networks to detect multiuser FSO-MIMO signals without channel estimation and reports simulated error rates close to maximum likelihood detection. Read it to see an early attempt at replacing channel estimation with learned detection in optical wireless links, and where that attempt currently falls short.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Multiuser interference model is unphysical: in FSO all users' signals superpose at every aperture, so the simulated multiuser case does not validate the central claim.","rationale":"The reader's weakest_assumption is exactly the concern I find most load-bearing: the multiuser interference scenario is modeled as one aperture carrying the target signal and others carrying interference, rather than as a physical superposition of all users' signals at every aperture. This is corroborated by the equations: Eq. (1) models a single user's signal at each aperture, and Eqs. (4)-(5) contain no inter-user interference term. Since the paper's title and abstract emphasize multiuser FSO-MIMO, the multiuser results are a central part of the claimed contribution. If the model is unphysical, Figs. 2c, 3c, 4c, 5c, and 6c do not demonstrate performance in a real multiuser interference channel. I considered whether the mislabeled 'blind' detector (trained with true labels) is a stronger concern; it is a semantic and practical issue, but the physical model flaw is more directly falsifiable and more fundamental to the multiuser results. The reader's reject verdict therefore stands. I also note the absence of code, data, and error bars, but the physical model issue alone justifies rejection. A concrete simulation test with the correct superposition model would settle whether the DNN's multiuser performance claim survives.","tokens_in":8981,"tokens_out":4647,"duration_ms":52795,"concrete_test":"Re-derive the received signal for N_u users all transmitting to N_r apertures on the same wavelength: y_i = R * sum_{l=1}^{N_u} sum_{j=1}^{N_t} I_{i,j,l} x_l + n_i, where I_{i,j,l} is the turbulence fading on the path from the j-th aperture of user l to the i-th receive aperture. Re-run the multiuser-interference simulations (Figs. 2c and 4c) with this superposition model while keeping the DNN architecture and training procedure unchanged. If the DNN's symbol error rate no longer tracks the perfect-CSI ML curve, the paper's central multiuser claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's multiuser interference model is described in Section II: 'one of the received N_r signals is the target signal and the others are interference signals.' This assumes the receive apertures carry different users' signals separately. In a physical FSO link, all users transmitting on the same wavelength are collected by every aperture, and the photodetector current at each aperture is a superposition of all users' intensity signals plus noise. Equations (1), (4), and (5) contain no inter-user interference term; they show only the target user's signal. Consequently, the curves labeled 'multiuser without resource allocation' simulate a spatially separated reception model, not multiuser FSO-MIMO interference. Because the title and abstract foreground multiuser scenarios, this assumption is load-bearing: if the received signal model is wrong, the multiuser performance claims (e.g., Fig. 4c) do not follow. This is not a disagreement with consensus but an internal inconsistency between the claimed physical scenario and the implemented equations. The reader's weakest_assumption identifies the same issue.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes deep neural network (DNN) based 'blind' detectors and a joint DNN detector-constellation shaping structure for free-space optical (FSO) MIMO links with intensity modulation and direct detection. It considers single-user and multiuser scenarios (with and without resource allocation), SISO and MIMO configurations, EGC and SC combining, and Gamma-Gamma turbulence from weak to strong. The main claim is that the DNN-based detectors, without channel state information, achieve symbol error rates close to, and for 4-QAM matching, maximum likelihood detection with perfect channel estimation, while allegedly having low complexity, cost, and latency. The results are presented as SER versus Es/N0 curves in Figs. 2 through 6.","tokens_in":9259,"tokens_out":5141,"duration_ms":51948,"significance":"If the results held, the paper would contribute a practical alternative to ML detection with perfect CSI in FSO-MIMO systems, particularly by avoiding explicit channel estimation. The claimed universality across turbulence regimes, combining schemes, and multiuser configurations is attractive. However, the significance is conditional on correcting a physically inconsistent multiuser signal model and on a fair treatment of the training labels used by the proposed 'blind' detector. The paper also ships no code and gives no statistical confidence measures, so the empirical claims are not yet verifiable. The combination of a wrong multiuser model and the unfair comparison with the ML baseline substantially weakens the contribution as it stands.","major_comments":[{"comment":"The multiuser signal model is physically inconsistent. In an FSO system with multiple users transmitting simultaneously on the same wavelength, the optical intensities from all users are collected by every receive aperture and the photodetector current is a superposition of all users' signals plus noise. Equation (1), however, contains only the target user's symbol x_l and has no summation over users; Eqs. (3)-(5) similarly describe only one user's signal. The description in Section II that 'one of the received N_r signals is the target signal and the others are interference signals' assumes spatial separation of users across receive apertures, which is neither a stated design condition nor the standard FSO multiuser channel. Because the title, abstract, and conclusion foreground multiuser interference, this modeling error is load-bearing: the curves labeled 'multiuser without resource allocation' in Figs. 2-6 simulate a spatially separated reception model, not multiuser FSO-MIMO interference. The physical scenario and the implemented equations are internally inconsistent.","section":"Section II, Eqs. (1)-(5)"},{"comment":"The claim that the proposed receiver is 'completely blind' is contradicted by the training procedure. Section III.A states that the one-hot vector is 'the label vector required for training the DNN,' meaning the DNN is trained with the true transmitted symbols. The conventional ML receiver is penalized for needing perfect or blind channel estimation, while the DNN is given ground-truth labels without accounting for how those labels are obtained in practice. If the labels require pilot transmission or a prior training phase, the comparison is unfair in terms of power, latency, and spectral efficiency. The paper should either explicitly state that the DNN is a supervised detector trained offline with known labels and compare against a conventional receiver with the same training overhead, or it should use a genuinely blind/unsupervised training objective. As written, the term 'blind' is misleading.","section":"Section III.A"},{"comment":"The simulation results are reported without any statistical confidence measures. No error bars, no train/test split, and no number of test symbols are given, so it is impossible to know whether the gap between the DNN curves and the ML curves in Figs. 2-6 is significant or merely a consequence of a particular random seed. Moreover, hyperparameters are tuned once on the SISO single-user case (Section IV) and reused for all other scenarios; Fig. 5 shows the gap grows with the number of apertures, which the authors attribute to this fixed tuning, but no retuning experiments are shown. The paper should provide confidence intervals and describe the data split, the number of Monte Carlo trials, and the test-set size.","section":"Section IV"},{"comment":"The complexity claim is asserted but never quantified. The abstract and conclusion state that the proposed structure has 'very few complexity' and 'greatly reduce[s] cost, and processing latency,' yet no analysis of the inference complexity (number of multiply-accumulate operations, number of parameters) or a comparison with the complexity of maximum likelihood detection is provided. The paper also ignores the cost of DNN training and the pilot overhead for obtaining labels. Without a concrete complexity model, the central practical motivation remains unsupported.","section":"Abstract and Section IV"}],"minor_comments":[{"comment":"The text says 'other DNN hyperparameters are exactly the same as descriptions of section III.B,' but it should refer to Section III.A; as written, it is self-referential and ambiguous.","section":"Section III.B"},{"comment":"Contribution 4 says 'SISO/MIMO (SC/ EEG) structures'; 'EEG' is a typo and should be 'EGC'.","section":"Section I, contribution list"},{"comment":"The Meijer-G function notation in Eq. (2) is incomplete; the arguments of G are not fully specified, which makes the distribution definition difficult to parse.","section":"Section II, Eq. (2)"},{"comment":"The conclusion says 'a few deferring number of apertures would greatly change the performance'; 'deferring' should be 'differing'.","section":"Section V"},{"comment":"The reference list contains an unusually high fraction of self-citations ([1]-[5], [17]-[18], and [22]), and [22] is an arXiv preprint whose title includes a nonstandard phrase 'under the shelf of Optical Communication'; the novelty claims in the Introduction should be supported by a broader independent literature review.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The multiuser physical model issue is the most serious concern; I would ask the authors to rework Section II and rerun the multiuser simulations with a correct superposition model, or to clearly restrict the claims to a single-user or spatially-separated-users scenario. The 'blind' labeling issue also needs a careful response: supervised training with true labels is acceptable if framed as an offline-trained detector, but then the comparison with ML+CSI must account for the training overhead. Please also check the citation pattern; the paper leans heavily on the author's own prior work, which makes it harder to assess the novelty claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — quick verdict on 1909.05147. The single-user part of this paper is genuinely interesting: the author shows a DNN can detect FSO signals without channel estimation and get close to ML with perfect CSI, and the end-to-end constellation shaping idea is a reasonable extension of work in fiber. The sweep over weak-to-strong turbulence and SISO/MIMO is useful. That part deserves to be seen.\n\nThe problem is the multiuser section. Section II says 'one of the received N_r signals is the target signal and the others are interference signals.' That is not how FSO works. All users' optical signals add at every aperture; the photocurrent at each aperture is a superposition of all users plus noise. Equations (1), (4), and (5) contain no interference term at all — they show only the target user's signal. So the curves labeled 'multiuser without resource allocation' simulate a physical scenario that does not exist. This is load-bearing because the title and abstract foreground multiuser. The multiuser claims collapse.\n\nOther soft spots: the receiver is called 'completely blind' but trained with true one-hot labels at the transmitter. That is standard supervised learning, not blind in the usual sense. No error bars, no train/test split, no code or data, so the simulation curves are not independently checkable. The complexity claim ('very few complexity') is asserted, not analyzed. And the reference list leans heavily on the author's own prior work, which doesn't help.\n\nWhere does that leave it? The single-user SISO/MIMO results are plausible and could be a useful contribution if reworked. The multiuser results should be redone with a proper superposition model or dropped. I'd send it to a serious referee, but with a clear request to fix the multiuser model and provide reproducibility details before it can be published. Right now, as written, the paper does not support its central claim.","headline":"The single-user DNN detector is plausible and worth a look, but the multiuser interference model is physically wrong, so the headline claims don't hold.","tokens_in":9704,"tokens_out":3420,"would_cite":false,"duration_ms":32979,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A neural receiver with no channel knowledge matches ideal maximum likelihood detection in FSO-MIMO links","keywords":["blind detection","deep neural network","free-space optical communication","FSO-MIMO","Gamma-Gamma turbulence","multiuser interference","constellation shaping","equal gain combining"],"falsifier":"Run the same 4-QAM and 16-QAM simulations with the physical superposition model—all users' turbulence-faded optical signals added at each receive aperture before detection—and compare the symbol error rate with the paper's one-target-aperture multiuser results; a material increase in SER would falsify the multiuser-interference claim as stated.","tokens_in":8809,"feed_emoji":"📡","tokens_out":5775,"duration_ms":56485,"temperature":0.7,"pith_summary":"This paper claims that a deep neural network can act as a fully blind detector in free-space optical MIMO links, with no channel estimation at all, and still come close to—and in the 4-QAM case match—maximum likelihood detection that assumes perfect channel state information. The claim covers single-user links, multiuser links with channel allocation, and multiuser links with interference, across weak to strong atmospheric turbulence and SISO/MIMO structures with equal-gain and selection combining. A second proposed structure adds a transmitter-side neural network that jointly shapes the constellation, and the paper argues this end-to-end design roughly doubles the improvement over the receiver-only detector. The reason a sympathetic reader would care is that the deep-learning detector removes the equalizer, pilots, and channel-estimation cost, so if the claim holds, practical FSO terminals could cut cost and latency while preserving near-optimal error rates.","feed_headline":"Blind neural detector matches ideal ML detection in FSO links","feed_subtitle":"Simulations from weak to strong turbulence show a receiver with no channel estimates can match maximum likelihood detection.","key_machinery":"The central object is a feed-forward deep neural network used as a blind detector: two input neurons take the real and imaginary parts of the combined received signal, hidden layers (four layers, forty neurons per layer, ReLU activations) transform it, and $M$ output neurons with softmax cross-entropy produce a one-hot estimate of the transmitted $M$-ary symbol. Training uses labeled one-hot vectors as ground truth, the Adam optimizer with learning rate $0.005$, and a batch size of $264$ over $1000$ iterations. In the joint structure a second DNN sits at the transmitter with $M$ input and two output neurons, and its complex output defines a learned constellation point, shaping the constellation so that turbulence has less effect; the receiver DNN then detects the shaped symbol. The paper's argument is carried by the ability of this pair to learn the nonlinear Gamma-Gamma channel behavior without explicit channel estimates.","core_discovery":"On its own terms, the paper's discovery is that the symbol error rate of a receiver that is given no channel state information—and that does not even estimate the channel blindly—is close to the symbol error rate of a maximum likelihood receiver that knows the turbulence fading exactly, provided the detector is a trained feed-forward deep network trained on labeled one-hot symbols. In the 16-QAM simulations the blind network tracks ML across weak, moderate, and strong Gamma-Gamma turbulence, and with 4-QAM the end-to-end network with transmitter-side constellation shaping matches the QAM-ML result for both SISO and MIMO. Because EGC and SC combining in FSO require no channel information, the comparison is held in the detection stage, and the paper attributes the residual gap to hyperparameters tuned once on the SISO single-user case and reused everywhere else.","pith_inferences":["The paper's multiuser-interference model assumes one receive aperture carries the target user while the other apertures carry interferers; in a real FSO link every user's optical field adds at every aperture, so the multiuser-interference results are conditional on a separation that physics does not provide.","A fair energy comparison would credit the blind structures with the power saved by not sending pilots for channel estimation; on a fixed total-energy budget the reported advantage over ML with perfect CSI could be larger than the figures show.","Because the authors state the hyperparameters were tuned once for the SISO single-user case, re-tuning per number of apertures and users is a direct testable extension that could close the MIMO and multiuser gaps.","The learned constellation shaping at the transmitter is a general idea: if it works under Gamma-Gamma turbulence, the same end-to-end training principle should transfer to other fading channels, though the paper only demonstrates FSO."],"forward_implications":["For 4-QAM, the end-to-end blind system reaches the symbol error rate of ideal QAM-ML detection in both SISO and MIMO, which implies the pilot overhead and channel-estimation stage can be removed without losing performance.","For 16-QAM, the blind detector stays close to QAM-ML over weak, moderate, and strong turbulence, so the claimed saving in complexity, cost, and latency does not require a benign channel.","Adding the transmitter-side constellation-shaping DNN improves performance substantially beyond the receiver-only DNN, with the paper attributing roughly half the end-to-end gain to shaping and half to neural detection.","Multiuser scheduling on the best channel can beat the single-user case, because the probability that all users simultaneously see bad turbulence falls as the number of users grows.","The gap between blind detection and perfect-CSI detection widens as the number of apertures or users grows, which the paper ties to hyperparameters tuned once for the SISO single-user scenario rather than per scenario."],"supporting_citations":[{"why":"Supplies the deep-learning geometric constellation shaping approach that the proposed DNN detector and joint shaping structure build on.","marker":"[6]"},{"why":"Defines the conventional blind channel-estimation receiver whose complexity the proposed fully blind detector removes.","marker":"[18]"},{"why":"Provides the Gamma-Gamma FSO-MIMO channel equations and EGC/ML receiver forms used as the conventional baseline.","marker":"[19]"},{"why":"Defines the cross-entropy loss function used to train the blind detector.","marker":"[20]"},{"why":"Provides the hyperparameter-tuning approach the paper says determines DNN performance across all scenarios.","marker":"[22]"}],"fun_headline_variants":["Blind neural detector rivals ML in FSO without channel info","Deep learning blind detection matches ML in FSO links","No channel estimates? Neural net matches ML in FSO","Blind DL receiver equals ML performance in FSO-MIMO","FSO-MIMO: Blind neural detection hits ML accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is the multiuser-interference model, in which one receive aperture is treated as carrying the desired user's signal and the other apertures as carrying interference; in physical FSO reception all users' optical signals add at every aperture, and if that physical summing replaces the assumed separation the multiuser claim may collapse.","fun_headline_variants_meta":{"raw":{"variants":["Blind neural detector rivals ML in FSO without channel info","Deep learning blind detection matches ML in FSO links","No channel estimates? Neural net matches ML in FSO","Blind DL receiver equals ML performance in FSO-MIMO","FSO-MIMO: Blind neural detection hits ML accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000219,"raw_usage":{"total_tokens":1429,"prompt_tokens":916,"completion_tokens":513,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":532,"completion_tokens_details":{"reasoning_tokens":431}},"tokens_in":532,"tokens_out":513,"duration_ms":5202,"temperature":1.0,"reasoning_tokens":431,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:14:04.658550+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same 4-QAM and 16-QAM simulations with the physical superposition model—all users' turbulence-faded optical signals added at each receive aperture before detection—and compare the symbol error rate with the paper's one-target-aperture multiuser results; a material increase in SER would falsify the multiuser-interference claim as stated.","supporting_citations":[{"cited_title":"A., & Vakili, V","cited_arxiv_id":null,"evidence_quote":"Defines the conventional blind channel-estimation receiver whose complexity the proposed fully blind detector removes."},{"cited_title":"R., & Ghassemlooy, Z","cited_arxiv_id":null,"evidence_quote":"Provides the Gamma-Gamma FSO-MIMO channel equations and EGC/ML receiver forms used as the conventional baseline."},{"cited_title":"Novel Suboptimal approaches for Hyperparameter Tuning of Deep Neural Network [under the shelf of Optical Communication]","cited_arxiv_id":"1907.00036","evidence_quote":"Provides the hyperparameter-tuning approach the paper says determines DNN performance across all scenarios."}],"review_version":1}