{"id":"7b94e21b-9586-4510-a970-1fdca583d079","arxiv_id":"2411.10101","paper_version":1,"verdict":"UNVERDICTED","confidence":"HIGH","novelty_score":1.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"A mini-review of the authors' prior work on VAE-based blind equalization, FPGA-implemented CNN equalizers, and spiking neural network equalizers for optical communications.","lead":"This paper surveys recent work by the authors on machine learning-based equalizers for optical communications. It is a compact entry point for engineers, but it presents no new results or data.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified; the VAE linear-channel assumption is explicitly acknowledged, and the CNN complexity claim is an empirical result from cited prior work.","rationale":"The reader's weakest_assumption about the VAE linear-channel assumption is a legitimate limitation but not a hidden premise: the paper explicitly states it and scopes Fig. 1 accordingly. A stronger candidate for scrutiny is the CNN complexity comparison, but that is an empirical result from referenced, presumably peer-reviewed work, and the preprint presents it with an explicit constraint. Given the paper's nature as a compact survey, the appropriate verdict remains UNVERDICTED rather than ACCEPT or REJECT. No internal inconsistency or unsupported assertion that would change the verdict was found.","tokens_in":4789,"tokens_out":6790,"duration_ms":62842,"concrete_test":"Worth running: reproduce the MAC-vs-BER Pareto analysis in Fig. 2 using the design-space exploration methodology/code from the cited FPGA work [11] and check whether the optimized CNN remains strictly better than the Volterra equalizer at the stated FPGA complexity limit; if the ordering changes, the 'CNN can outperform Volterra' claim should be recast as setup-specific.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper is a review summarizing the authors' prior work. The strongest concrete assertions are (i) a VAE-based equalizer can outperform CMA for PCS-QAM on a linear dispersive channel (Fig. 1), and (ii) an optimized CNN can outperform linear and Volterra equalizers under a MAC constraint in a specific IM/DD setup (Fig. 2). Both are scoped claims. The linear-channel/AWGN assumption underlying the VAE cost function is stated explicitly in Section 2: 'Unfortunately, the mathematical derivation... relies on the assumption of a linear channel and additive white Gaussian noise.' Fig. 1's caption confirms the simulation is for a linear dispersive channel, and the paper does not extend this result to nonlinear channels, instead citing vector-quantized VAE [9,10]. The CNN comparison is presented as an empirical design-space exploration with a stated FPGA complexity limit, with details in [11,12]. While the preprint itself contains no new derivations or data, the claims it makes are internally consistent and appropriately attributed. No load-bearing concern that would invalidate the review's message is identified.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper is a short review that presents recent advances in machine-learning-aided digital signal processing for optical communications, based on the authors' prior work. It covers three areas: (i) a variational-autoencoder (VAE) derived blind cost function for equalizers, shown to outperform CMA for probabilistically shaped QAM in a linear dispersive channel; (ii) a CNN-based equalizer for short-reach IM/DD systems that can outperform linear and Volterra equalizers under a MAC-operation constraint, with FPGA demonstrations; and (iii) energy-efficient equalizer implementations using spiking neural networks on neuromorphic hardware. The paper contains no new derivations or experimental data; it summarizes results that are fully referenced to the authors' previous publications, and it explicitly discusses the applicability limits of the VAE approach.","tokens_in":4957,"tokens_out":5111,"duration_ms":50502,"significance":"If the presented results are correct, the paper provides a useful and accessible summary of a coherent line of research showing that ML-based equalizers are not inherently impractical for high-speed optical transceivers. Its main strengths are the explicit scoping of the key claims: the VAE result is clearly limited to a linear channel with AWGN, and the CNN complexity comparison is stated under a specific MAC-operations constraint on a given FPGA platform. The paper is honest about these limitations and points to further work (e.g., vector-quantized VAEs for nonlinear channels). As a review, it does not establish new results, but it may serve as a valuable introduction for readers and demonstrates that the authors' claims are reproducible through the cited sources.","major_comments":[],"minor_comments":[{"comment":"The opening sentence 'It is a common misconception that ML-based equalizers are more computationally demanding than conventional equalizers' is too absolute; the subsequent sentences show that this statement holds only for a carefully optimized CNN in a specific IM/DD scenario under a MAC constraint. I recommend rewording to something like 'It is sometimes assumed that ML-based equalizers are necessarily more computationally demanding than conventional equalizers, but this is not always the case' to avoid giving the impression that ML equalizers are generally competitive in complexity.","section":"Section 3, first paragraph"},{"comment":"The terms A(phi,y) and B(theta,phi,y) appear in the block diagram but are not defined or explained in the text. Please add a brief explanation of these quantities in the caption or in the text.","section":"Section 2, Fig. 1 right panel"},{"comment":"The caption mentions a red dashed vertical line indicating the FPGA complexity limit, but the text does not discuss what 'feasible' and 'too complex' mean in terms of the MAC budget or the FPGA utilization. A single sentence in the text clarifying this would help the reader interpret the Pareto front.","section":"Section 3, Fig. 2 caption"},{"comment":"The section on spiking neural networks is very brief and does not discuss the trade-off between spike encoding and equalization performance, nor does it give a comparison with the earlier CNN results. Adding one or two sentences summarizing the key advantage and the main challenge of SNN equalizers would make the section more balanced.","section":"Section 4"},{"comment":"The abstract is almost identical to the first sentence of the introduction; consider making the abstract more informative, for example by mentioning the two key concrete results (VAE outperforming CMA for PCS-QAM and CNN outperforming Volterra under a MAC constraint).","section":"Abstract and Introduction"}],"recommendation":"minor_revision","confidential_remarks":"The paper is essentially a summary of the authors' own previous publications, not a broad review of the field despite the title 'Recent Advances on Machine Learning-aided DSP'. This is not a fatal flaw if the journal expects an invited or concise overview, but the editor may want to consider whether the scope matches the title. The technical claims are appropriately attributed and scoped, and I found no internal inconsistencies or unsupported load-bearing statements."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague—\n\nThis is a short review of the authors' own prior work on ML-based equalizers for optical communications. If you know the key papers by this group, there is nothing new here: no new algorithm, derivation, or measurement. But the paper is honest, clearly written, and useful as a compact orientation to their line of research.\n\nWhat it does well: the VAE-based blind equalizer result is carefully scoped. In Section 2 they explicitly state that the derivation assumes a linear channel and AWGN, and they cite vector-quantized VAEs for the nonlinear case. The CNN/FPGA complexity comparison is presented as an empirical design-space exploration from earlier papers, not as a universal claim. The figures are clean, and the references point to the underlying published work.\n\nThe soft spots are mostly about scope. The title says 'Recent Advances' but the content is largely a summary of the authors' own contributions; external literature appears only in passing. The claim that ML equalizers are not necessarily more computationally demanding is stated as a 'common misconception' and supported only by one specific IM/DD scenario with a hard MAC constraint. That is a valid demonstration, but it is not a general result, and the paper does not try to overclaim beyond the cited work. No code or data are shipped, which is fine for a review but means there is nothing independently reproducible here.\n\nWho is this for? A newcomer who wants a quick map of this group's work on VAE, CNN, and SNN equalizers, or a workshop audience. A researcher looking for new results will be disappointed. If this crossed my desk as a journal submission, I would desk-reject it for lack of new content; if it is an invited summary or a conference digest, it is acceptable as is.\n\nI would not send this to full peer review as a research paper, but it does not contain a load-bearing flaw. The claims are scoped and attributed. My recommendation: if the venue expects original research, reject; if it explicitly welcomes reviews or summaries, accept with light copyediting.","headline":"Self-review of the authors' own ML equalizer work: honest and clear, but no new results and too narrow for the title.","tokens_in":5445,"tokens_out":2955,"would_cite":false,"duration_ms":29727,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"For probabilistically shaped coherent QAM, a variational-autoencoder blind equalizer outperforms CMA, and for short-reach IM/DD at 40 GBd, a complexity-optimized CNN outperforms linear and Volterra equalizers under a fixed MAC budget.","keywords":["machine learning equalization","optical communications","variational autoencoder","blind equalization","convolutional neural network","FPGA implementation","spiking neural network","intensity-modulation direct detection"],"falsifier":"Transmit probabilistically shaped 64-QAM over a dispersive coherent link with strong Kerr nonlinearity at the symbol rates of Figure 1, and compare the variational-autoencoder equalizer with CMA; if the VAE equalizer no longer reaches a lower symbol-error rate than CMA, the paper's headline comparison holds only for linear channels and its nonlinear-channel promise rests on an unreported vector-quantized variant.","tokens_in":4603,"feed_emoji":"📶","tokens_out":10725,"duration_ms":95036,"temperature":0.7,"pith_summary":"The paper aims to establish that machine-learning-based equalizers are ready to compete with, and in constrained settings outperform, classical DSP in optical communications. It reports two main results: a variational-autoencoder-derived blind cost function that beats a CMA-based equalizer for probabilistically shaped 64-QAM on a coherent link, and a convolutional-neural-network equalizer whose structure is optimized under a fixed multiply-accumulate budget that beats linear and Volterra equalizers in a short-reach IM/DD link at 40 GBd. It also surveys spiking-neural-network equalizers as an energy-efficient alternative, including a digital emulation that reportedly uses less energy than a classical digital equalizer. A sympathetic reader would take the paper as evidence that ML equalizers can be practical, provided the cost function matches the channel or the hardware budget is part of the design.","feed_headline":"Blind VAE beats CMA; optimized CNN beats Volterra on FPGA","feed_subtitle":"A blind variational-autoencoder equalizer and a hardware-optimized CNN beat classical digital signal processing.","key_machinery":"The machinery that carries the first claim is the variational-autoencoder blind cost function, a lower bound on the log-likelihood of received symbols that combines a KL-divergence term between the soft-demapper posterior $q$ and the source distribution $P(\\mathbf x)$ with an expected squared-error term $\\mathbb{E}_q[\\|\\mathbf y - \\mathbf h_\\theta * \\mathbf x\\|^2]$ involving an estimated channel impulse response $\\mathbf h_\\theta$. This joint estimator-equalizer-demapper objective lets the demapper's soft outputs steer the equalizer taps without pilots. The machinery that carries the second claim is a design-space exploration in MAC operations per symbol: for each candidate CNN architecture the BER is measured against the number of multiply-accumulate operations, and the Pareto front is compared with linear and Volterra equalizers at the FPGA's complexity limit.","core_discovery":"On the paper's own terms, the central discoveries are two. First, replacing the usual mean-squared-error or constant-modulus cost with a variational-autoencoder cost function yields a blind equalizer that adapts using only received noisy symbols and source statistics; in simulation over a linear dispersive dual-polarization channel with probabilistically shaped 64-QAM, this VAE-based linear equalizer reaches lower symbol-error rates than CMA, which the paper reports as not converging. Second, a CNN equalizer with a carefully chosen layer configuration outperforms both linear and nonlinear Volterra equalizers in a short-reach IM/DD scenario once the number of multiply-accumulate operations per symbol is constrained by a real-time FPGA implementation; the paper shows this in a design-space exploration at 40 GBd. The paper additionally reports that a trainable NN equalizer including real-time training fits on an FPGA up to 20 GBd. The paper is explicit that the VAE cost derivation assumes a linear channel with AWGN, and for nonlinear channels it points to a vector-quantized variant without presenting results in this preprint.","pith_inferences":["Editorial inference: if the VAE cost function is replaced by its vector-quantized variant, the same blind-equalization principle may extend to strongly nonlinear channels such as short-reach passive optical networks with amplifier saturation; this is a testable prediction the preprint leaves open.","Editorial inference: the operation-count-normalized comparison used for the CNN result suggests a general benchmark for DSP: report BER against MAC operations per symbol, so linear, Volterra, CNN, and spiking equalizers can be compared on equal footing.","Editorial inference: the success of the complexity-constrained CNN suggests the same algorithm-hardware co-design could be applied to other receiver blocks, such as carrier phase estimation or demapping, without requiring new theory."],"forward_implications":["A VAE-based blind equalizer makes probabilistically shaped QAM usable in the startup phase of coherent links, where CMA-based equalizers fail to converge.","The same VAE objective provides a channel impulse-response estimate at no extra cost, which the paper suggests could be used for integrated communications and sensing.","A CNN equalizer can be real-time implementable at 40 GBd on an FPGA while outperforming Volterra equalizers, so ML equalization is not inherently too complex for short-reach systems.","Real-time training of an NN equalizer on FPGA is feasible up to 20 GBd, enabling adaptive ML equalization in deployed transceivers.","Neuromorphic spiking-neural-network equalizers, and even digital emulations of them, are a plausible route to lower energy per bit than classical DSP."],"supporting_citations":[{"why":"It supplies the VAE-based blind equalization and channel-estimation method and the simulation results comparing VAE-LE with CMA for probabilistically shaped 64-QAM.","marker":"[6]"},{"why":"It provides the prior unsupervised VAE equalization framework that the paper's cost function extends to a blind linear equalizer.","marker":"[7]"},{"why":"It supports the claim that VAE-based equalizers improve the startup and bootstrap phase of blind equalization.","marker":"[8]"},{"why":"It reports the real-time FPGA demonstrator of an ANN equalizer that grounds the hardware-feasibility claim at 40 GBd.","marker":"[12]"},{"why":"It supplies the FPGA architecture and complexity analysis for CNN-based equalization with gigabit throughput.","marker":"[11]"},{"why":"It provides fully-blind NN equalization results in passive optical networks with strong nonlinear distortions.","marker":"[14]"},{"why":"It introduces the energy-efficient spiking-neural-network equalizer for IM/DD systems.","marker":"[16]"},{"why":"It shows that a digital emulation of an SNN-based equalizer can be more energy-efficient than a classical digital equalizer.","marker":"[19]"}],"fun_headline_variants":["Blind VAE equalizer outsets CMA; FPGA-tuned CNN tops Volterra","VAE blind beats CMA; CNN beats Volterra on FPGA","ML DSP: blind VAE beats CMA, CNN beats Volterra","Blind VAE and FPGA CNN beat classical equalizers","Machine-learning equalizers: VAE blind, CNN beats Volterra"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The variational-autoencoder equalizer's advantage over CMA is proven only under the assumption that the optical channel is linear and the only noise is ordinary additive white Gaussian noise; for the nonlinear distortions common in short-reach links the paper points to a different variant but gives no supporting measurements.","fun_headline_variants_meta":{"raw":{"variants":["Blind VAE equalizer outsets CMA; FPGA-tuned CNN tops Volterra","VAE blind beats CMA; CNN beats Volterra on FPGA","ML DSP: blind VAE beats CMA, CNN beats Volterra","Blind VAE and FPGA CNN beat classical equalizers","Machine-learning equalizers: VAE blind, CNN beats Volterra"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000434,"raw_usage":{"total_tokens":2127,"prompt_tokens":778,"completion_tokens":1349,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":394,"completion_tokens_details":{"reasoning_tokens":1257}},"tokens_in":394,"tokens_out":1349,"duration_ms":12313,"temperature":1.0,"reasoning_tokens":1257,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:57:36.625237+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Transmit probabilistically shaped 64-QAM over a dispersive coherent link with strong Kerr nonlinearity at the symbol rates of Figure 1, and compare the variational-autoencoder equalizer with CMA; if the VAE equalizer no longer reaches a lower symbol-error rate than CMA, the paper's headline comparison holds only for linear channels and its nonlinear-channel promise rests on an unreported vector-quantized variant.","supporting_citations":[{"cited_title":"Blind Equalization and Channel Estimation in Coherent Optical Communications Using Variational Autoencoders","cited_arxiv_id":"2204.11776","evidence_quote":"It supplies the VAE-based blind equalization and channel-estimation method and the simulation results comparing VAE-LE with CMA for probabilistically shaped 64-QAM."},{"cited_title":"Unsupervised linear and nonlinear channel equalization and decoding using variational autoencoders,","cited_arxiv_id":null,"evidence_quote":"It provides the prior unsupervised VAE equalization framework that the paper's cost function extends to a blind linear equalizer."},{"cited_title":"Improving the Bootstrap of Blind Equalizers with Variational Autoencoders","cited_arxiv_id":"2301.06576","evidence_quote":"It supports the claim that VAE-based equalizers improve the startup and bootstrap phase of blind equalization."},{"cited_title":"Real-Time FPGA Demonstrator of ANN-Based Equalization for Optical Communications","cited_arxiv_id":"2402.15288","evidence_quote":"It reports the real-time FPGA demonstrator of an ANN equalizer that grounds the hardware-feasibility claim at 40 GBd."},{"cited_title":"CNN-Based Equalization for Communications: Achieving Gigabit Throughput with a Flexible FPGA Hardware Architecture","cited_arxiv_id":"2405.02323","evidence_quote":"It supplies the FPGA architecture and complexity analysis for CNN-based equalization with gigabit throughput."},{"cited_title":"Fully-blind Neural Network Based Equalization for Severe Nonlinear Distortions in 112 Gbit/s Passive Optical Networks","cited_arxiv_id":"2401.09579","evidence_quote":"It provides fully-blind NN equalization results in passive optical networks with strong nonlinear distortions."},{"cited_title":"Energy-efficient spiking neural network equalization for IM/DD systems with optimized neural encoding,","cited_arxiv_id":null,"evidence_quote":"It introduces the energy-efficient spiking-neural-network equalizer for IM/DD systems."},{"cited_title":"Efficient FPGA Implementation of an Optimized SNN-based DFE for Optical Communications","cited_arxiv_id":"2409.08698","evidence_quote":"It shows that a digital emulation of an SNN-based equalizer can be more energy-efficient than a classical digital equalizer."}],"review_version":1}