{"id":"5a0b66c6-06be-4a3c-bac7-83d3ef09af35","arxiv_id":"2506.19707","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A Gaussian boson sampling device is used as a random feature generator for extreme learning machines, reaching 95.86% on MNIST and 85.95% on Fashion-MNIST.","lead":"Researchers used a large photonic quantum device called Jiuzhang to create random features and classify handwritten digits and fashion items, reaching about 95.86% accuracy on MNIST and 85.95% on Fashion-MNIST. The work tests whether a quantum sampling device can be useful for machine learning, though the accuracies are below what ordinary classical neural networks already achieve.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The GBS feature map uses only 16-mode marginal click statistics per group, which are classically simulable; without a classical simulation of the same map, the claim that GBS enhances recognition is not established.","rationale":"The reader's verdict is CONDITIONAL, and this stress-test agrees. The strongest claim—that GBS-based ELM/RVFL beats the listed baselines—is not internally contradicted, and the empirical accuracies are plausible. However, the most significant gap is that the feature map uses only marginal statistics on small mode subsets, which is classically simulable, and the paper provides no classical simulation of the same map. This does not invalidate the measured accuracies, but it undermines the interpretation that GBS is the enabling resource. The missing simulation is concrete, testable, and directly addresses whether the central claim has quantum significance. The paper's own comparison to coherent-state input is insufficient because it changes the feature map. A revised version should add this classical baseline, stronger classical ELM baselines, and error bars. These changes would not necessarily change the accuracy numbers but would determine whether the 'GBS enhances image recognition' conclusion survives. Therefore the verdict should remain CONDITIONAL: the paper is acceptable contingent on providing the classical simulation and stronger baselines.","tokens_in":10529,"tokens_out":4633,"duration_ms":55186,"concrete_test":"Implement the same feature map classically. For each image, apply the same PCA and scaling, select the same 16-mode groups and temporal modes, and compute the exact probabilities of the selected N computational bases using the 16-mode Gaussian marginal (hafnian/Torontonian) derived from the specified covariance matrix and loss parameters (as described for Jiuzhang 4.0 in Refs. [7,8]). Use these probabilities (or 5 million classically sampled outcomes from this marginal) as the GELM/GRVFL input, train the same pseudoinverse classifier, and compare test accuracy on the same split. If the accuracy matches 95.86% and 85.95% within the Monte Carlo error, the GBS device provides no advantage.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central empirical claim is that the GBS-based GELM/GRVFL classifiers outperform the listed baselines, with GBS as the enabling nonlinear transformation. However, the feature vector is constructed by counting outcomes on only the 16 selected modes per group (Section 'Third, as shown in FIG. 1(a)...' and 'Fourth, we count the number of samples on the computational bases for the selected modes'). These counts are frequencies from the marginal photon-number distribution of a Gaussian state restricted to those 16 modes. For any Gaussian state, this marginal is fully characterized by the 16x16 covariance submatrix, and sampling from it or computing exact probabilities of the 2^16 possible patterns is polynomial-time classically feasible (e.g., via hafnian or conditional Gaussian sampling). The paper never compares against a classical simulation of this same feature map. The only 'classical counterpart' tested is coherent-state input, which produces a different marginal distribution, so it does not isolate whether the squeezed-state feature map itself requires quantum resources. Consequently, the observed accuracies may be entirely reproducible with a classical computer, and the paper's statement that 'GBS is a better choice than coherent states in transforming data for image classification' does not establish that GBS (as opposed to the specific marginal statistics) is responsible for the improvement. This is load-bearing because the scientific significance of the result depends on GBS contributing something non-classical; without a classical-simulation baseline, the headline claim reduces to an implementation detail. The reader's weakest_assumption correctly identifies this gap, and the paper's own limitation—omitting any classical simulation of the selected-mode marginals—is not discussed in the text.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript reports a scheme for image classification using Gaussian boson sampling (GBS) as the nonlinear feature extractor in an extreme learning machine (ELM) and random vector functional-link (RVFL) classifier. Input images are reduced by PCA to M features, each feature selects a spatial-temporal mode pair in an 8176-mode GBS device, and the photon-count statistics on the selected computational bases form the feature vector. The output layer is trained analytically by pseudo-inverse. On MNIST and Fashion-MNIST, the authors report testing accuracies of 95.86% and 85.95% for GRVFL, outperforming linear-kernel SVC and three physical ELM implementations. They also study the effect of the number of computational bases, the number of PCA features, the number of GBS samples, and the device efficiency.","tokens_in":10768,"tokens_out":6515,"duration_ms":68427,"significance":"The experimental execution is a strength: a large-scale photonic device is used in a concrete machine-learning pipeline with standard benchmarks and reproducible analytical training. The hyperparameter and efficiency studies are useful. However, the scientific significance depends on whether GBS provides an advantage over classically available nonlinear feature maps, and this is not established. The feature map uses only the marginal photon-number distribution on a small number of modes, which is classically simulable for Gaussian input states. Without a classical simulation of the same feature map, the results are best viewed as a demonstration of a GBS-implemented feature extractor, not as evidence of quantum-enhanced recognition.","major_comments":[{"comment":"The feature vector is constructed by counting GBS samples on the computational bases of the selected 16 modes per group (32 modes for M=32). For a Gaussian input state and a known linear interferometer, the marginal photon-number distribution on these modes is fully characterized by the corresponding covariance submatrix; sampling from this marginal or computing the 2^m pattern probabilities (m ≤ 32) is classically efficient. The paper does not compare against a classical simulation of this exact feature map. The coherent-state comparison in Table I uses a different marginal distribution, so it cannot isolate whether the squeezed-state feature map itself has any quantum advantage. The authors should either implement a classical simulation of the same feature map (for example, by sampling from the Gaussian marginal or using their MPS samples) and compare accuracies, or restrict the claims to the statement that a GBS device can implement this feature map. As written, the claim that 'GBS is a better choice than coherent states' does not support the title's implication that GBS is the enabling resource.","section":"Fourth step (computational basis counting) and Table I"},{"comment":"The headline accuracies in Table I are reported without error bars or confidence intervals, although the hyperparameter study (Fig. 3) reports double standard error for the same models. The reported margins over the linear-kernel SVC baseline (1.56% on MNIST, 2.05% on Fashion-MNIST) may be sensitive to the random selection of computational bases and to the particular 5-million-sample subset. Additionally, linear-kernel SVC is a weak baseline for these datasets; without stronger classical baselines (for example, RBF-kernel SVM, random forest, or a classical ELM with random features of comparable dimension), the statement that the results 'surpass classical methods' is not well supported.","section":"Table I and Fig. 2"}],"minor_comments":[{"comment":"The text states that 9 million samples are generated initially, while the main results and hyperparameter default use 5 million samples; the relationship between these two numbers should be clarified explicitly.","section":"First step and Fig. 3(c,g)"},{"comment":"Reference [7] contains the informal sentence 'this article will be uploaded to arXiv soon,' which should be replaced with a proper citation or removed.","section":"Reference [7]"},{"comment":"The description of how PCA feature values are mapped to temporal mode indices would benefit from a concrete example or pseudo-code, as the current text is ambiguous about the grouping and partition assignment.","section":"Third step (mode mapping description)"},{"comment":"The abbreviations 'GELM' and 'GRVFL' are used without explicit expansion in the text; they should be defined at first occurrence for readers outside the quantum information community.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is from a group with an outstanding experimental record, and the experimental effort is substantial. The key issue is interpretational: the authors should either provide a classical simulation baseline for the same feature map or substantially soften the claim that GBS provides an enhancement. I recommend major revision on these grounds."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read on arXiv:2506.19707. What's actually new: this is the largest GBS device used in a machine-learning pipeline to date, with an 8176-mode temporal-spatial interferometer and about 2200 average clicks. The encoding trick—mapping PCA features to temporal modes and using click frequencies on selected computational bases as the ELM feature map—is a clever way to exploit the device's structure without needing programmability. The accuracy numbers (95.86% on MNIST, 85.95% on Fashion-MNIST) are plausible, the hyperparameter study is careful, and the experimental protocol is described in enough detail that someone could reproduce it. That is real value.\n\nThe stress-test note is on target. The feature vector per group is built from photon-number statistics on only 16 selected modes. For a Gaussian state, that marginal distribution is classically simulable in polynomial time. The coherent-state comparison is not the right classical baseline—it produces a different distribution, so it doesn't isolate whether the squeezed-state feature map itself is doing something nonclassical. Without a classical simulation of the same feature map, the conclusion 'GBS is a better choice than coherent states' is not established. That's a load-bearing gap, not a nitpick.\n\nOther soft spots, in proportion: the baselines are weak. Linear SVC at 92.9% on MNIST is a low bar; a simple RBF SVM on PCA features likely clears 98%. So the headline 'surpasses classical method SVC with linear kernel' is technically true but not impressive. No error bars on the headline accuracies, though the cross-validation plots do show them. No code or data release, which matters for benchmarking claims. These are all fixable.\n\nWhat holds up: the demonstration that a large GBS device can be integrated into a stochastic neural network and produce reasonable classification accuracy on standard benchmarks. The engineering and the experimental results are real. The paper is not circular or self-referential in a problematic way—the protocol is straightforward and the numbers are not fitted to the test set.\n\nOverall: yes, this deserves a serious referee. The experimental scale is notable and the method is clean enough to engage with. But the referee should require a classical simulation of the same feature map, stronger classical baselines, and error bars on the central numbers. As it stands, the paper is a solid experimental report with an overreaching quantum-significance claim.\n\nRecommendation: send it to peer review, but expect heavy revision. The right audience is people working on quantum reservoir computing and photonic ML—they'll want to see this, and they'll also want to see the classical simulation they're missing.","headline":"A genuinely large-scale GBS device wired into an ELM for MNIST is worth a serious look, but the paper overclaims what GBS uniquely buys you, since the feature map rests on classically simulable 16-mode marginals.","tokens_in":11456,"tokens_out":2118,"would_cite":false,"duration_ms":25340,"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":"A GBS-based classifier on an 8176-mode photonic processor reaches 95.86% accuracy on MNIST and 85.95% on Fashion-MNIST, beating linear-kernel SVC and prior physical ELM experiments.","keywords":["Gaussian boson sampling","extreme learning machine","random vector functional-link","image classification","MNIST","Fashion-MNIST","photonic processor","quantum machine learning"],"falsifier":"Take the same PCA preprocessing, mode-selection rule, computational-basis selection, and linear readout, but replace the GBS device with a classical sampler that draws from the 16-mode marginal photon-number distribution of the same Gaussian state (or computes those probabilities exactly); if the accuracy reaches or exceeds 95.86% on MNIST and 85.95% on Fashion-MNIST, the claim that GBS is the enabling resource is falsified.","tokens_in":10283,"feed_emoji":"⚛️","tokens_out":7544,"duration_ms":71125,"temperature":0.7,"pith_summary":"This paper proposes treating a Gaussian boson sampling device as a random nonlinear feature layer for image classification, following the extreme learning machine recipe: fix the quantum device as the hidden layer and train only a linear readout by pseudo-inverse. Using the 8176-mode Jiuzhang processor, the authors map PCA-compressed MNIST and Fashion-MNIST images onto mode selections, count GBS samples on the selected computational bases, and feed those counts to a linear classifier. They report testing accuracies of 95.86% on MNIST and 85.95% on Fashion-MNIST for the RVFL variant, beating linear-kernel SVC and three previous physical ELM experiments, and show that squeezed-state GBS features outperform coherent-state versions of the same feature map. The point of the paper is that a large-scale GBS device can serve as a practical, minimally programmable feature extractor for real-world machine learning tasks.","feed_headline":"Photonic GBS classifier beats SVM on MNIST and Fashion-MNIST","feed_subtitle":"Random GBS features beat linear SVM and prior optical ELM experiments on both benchmarks.","key_machinery":"The central object is a GBS-based feature map: a fixed random nonlinear transformation in which image-derived PCA features select spatial-temporal modes of an 8176-mode interferometer, and the empirical photon-number counts on selected computational bases become the feature vector for a pseudo-inverse-trained linear classifier. ELM uses only these transformed features, while RVFL concatenates the original data with the transformed features. The scheme's key property is that programmability is reduced to choosing which modes to count, while the expensive nonlinearity is supplied by the quantum device.","core_discovery":"The central claim is that GBS can be used as the hidden layer of an ELM/RVFL classifier and that this yields better accuracy than a classical linear SVM and prior optical ELM baselines on standard benchmarks. The classifier's only trainable part is a linear readout; the nonlinear transformation is produced by mapping each PCA feature to a spatial-temporal mode pair in the 8176-mode interferometer, counting samples on selected computational bases, and using renormalized counts as features. The authors demonstrate this on Jiuzhang, with approximately 2200 average photon clicks, and report GRVFL accuracies of 95.86% on MNIST and 85.95% on Fashion-MNIST. They further show that coherent-state input performs worse, that accuracy improves with efficiency and sample count, and that combining GBS features with original data beats using either alone.","pith_inferences":["The feature vector is built from the marginal photon-number statistics on just 16 of 8176 modes per group, and the input Gaussian state is fully characterized; a classical program can therefore sample or compute those marginals directly. Comparing the same feature map fed by classical samples would test whether the quantum device is the actual source of the accuracy gain, a comparison the paper do","The hyperparameter trends suggest the scheme saturates around 5 million samples; if that saturation is generic, cheaper or faster samplers could reproduce the features, and the practical bottleneck shifts from sampling to the classical readout training.","The same mode-selection encoding could be applied to other stochastic photonic devices, and the 16-mode group structure leaves room for larger groups or multiple selections per image, which may raise accuracy without retraining the device."],"forward_implications":["A GBS device can act as a fixed random feature layer for image classification, with training reduced to a pseudo-inverse readout and no iterative optimization.","The GRVFL variant, which concatenates GBS features with raw pixels, reaches 95.86% on MNIST and 85.95% on Fashion-MNIST, above the linear-kernel SVC baselines of 92.9% and 83.9%.","Squeezed-state GBS features outperform coherent-state features with the same mean photon number, indicating that the non-classical photon statistics contribute to the transformation.","Classification accuracy increases with GBS system efficiency and with the number of samples, so hardware improvements should translate directly into better machine-learning performance.","Because the same mode-selection and counting scheme requires little programmability, the approach can scale to larger mode counts without redesigning the readout."],"supporting_citations":[{"why":"Supplies the 8176-mode temporal-spatial photonic processor and the experimental GBS samples.","marker":"[7]"},{"why":"Defines the extreme learning machine architecture and the pseudo-inverse training rule used for the readout.","marker":"[24]"},{"why":"Defines the random vector functional-link network that motivates concatenating original and transformed features.","marker":"[25]"},{"why":"Provides the MNIST dataset used for the 95.86% accuracy result.","marker":"[26]"},{"why":"Provides the Fashion-MNIST dataset and the linear-kernel SVC baseline (92.9%/83.9%) that the results are compared against.","marker":"[27]"},{"why":"A large-scale GBS-based reservoir computing ELM experiment that the MNIST result is compared with.","marker":"[31]"},{"why":"A hot-atomic-vapor optical ELM whose MNIST accuracy the paper claims to surpass.","marker":"[32]"},{"why":"A free-space optical ELM baseline on MNIST that the paper claims to surpass.","marker":"[33]"},{"why":"The matrix-product-state classical simulation method used to generate samples at different efficiencies for the robustness study.","marker":"[8]"},{"why":"The broad learning system principle cited to explain why combining GBS features with original data helps.","marker":"[36]"}],"fun_headline_variants":["Jiuzhang GBS classifier beats SVM on image benchmarks","Gaussian boson sampling tops linear SVM on MNIST","GBS outclasses SVM on MNIST and Fashion-MNIST","Quantum GBS enhances image recognition over classical"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The accuracy claim rests on the assumption that the photon-count statistics from the selected modes are a useful feature representation that a classical computer cannot practically reproduce, a comparison the paper does not make.","fun_headline_variants_meta":{"raw":{"variants":["Jiuzhang GBS classifier beats SVM on image benchmarks","Gaussian boson sampling tops linear SVM on MNIST","GBS outclasses SVM on MNIST and Fashion-MNIST","Quantum GBS enhances image recognition over classical"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000683,"raw_usage":{"total_tokens":3088,"prompt_tokens":924,"completion_tokens":2164,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":540,"completion_tokens_details":{"reasoning_tokens":2098}},"tokens_in":540,"tokens_out":2164,"duration_ms":14973,"temperature":1.0,"reasoning_tokens":2098,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:27:20.720278+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the same PCA preprocessing, mode-selection rule, computational-basis selection, and linear readout, but replace the GBS device with a classical sampler that draws from the 16-mode marginal photon-number distribution of the same Gaussian state (or computes those probabilities exactly); if the accuracy reaches or exceeds 95.86% on MNIST and 85.95% on Fashion-MNIST, the claim that GBS is the enabling resource is falsified.","supporting_citations":[{"cited_title":"Zhong, H","cited_arxiv_id":null,"evidence_quote":"Supplies the 8176-mode temporal-spatial photonic processor and the experimental GBS samples."},{"cited_title":"Fujiyoshi, T","cited_arxiv_id":null,"evidence_quote":"Defines the extreme learning machine architecture and the pseudo-inverse training rule used for the readout."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the random vector functional-link network that motivates concatenating original and transformed features."},{"cited_title":"Pedregosa, G","cited_arxiv_id":null,"evidence_quote":"Provides the MNIST dataset used for the 95.86% accuracy result."},{"cited_title":"Lecun, L","cited_arxiv_id":null,"evidence_quote":"Provides the Fashion-MNIST dataset and the linear-kernel SVC baseline (92.9%/83.9%) that the results are compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A hot-atomic-vapor optical ELM whose MNIST accuracy the paper claims to surpass."},{"cited_title":"Hotelling, Analysis of a complex of statistical vari- ables into principal components, Journal of Educational Psychology 24, 417 (1933)","cited_arxiv_id":null,"evidence_quote":"A free-space optical ELM baseline on MNIST that the paper claims to surpass."},{"cited_title":"Zhong, Y.-H","cited_arxiv_id":null,"evidence_quote":"The matrix-product-state classical simulation method used to generate samples at different efficiencies for the robustness study."},{"cited_title":"Azam and R","cited_arxiv_id":null,"evidence_quote":"The broad learning system principle cited to explain why combining GBS features with original data helps."}],"review_version":2}