{"id":"1db61ac5-d7f5-4a41-a082-9717945ed19a","arxiv_id":"2505.11238","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Coincidence-based QELMs outperform intensity-based ones on a 0/1 MNIST task, and simulations predict that indistinguishable photons give a growing expressivity advantage as photon number increases, though the 2-photon experiment shows no accuracy advantage.","lead":"A research team built a quantum extreme learning machine from pairs of photons scrambled by a multimode fiber and showed that measuring photon coincidences beats measuring only intensities on a small digit-classification task. The interesting open part is the authors' simulation-based claim that using indistinguishable photons raises the expressivity of the machine faster as more photons are added, which points to a possible path toward scalable photonic machine learning.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim rests on feature-matrix rank as a proxy for expressivity, yet the paper's own experiment contradicts that proxy: at 22 detectors the IELM has higher rank but lower accuracy than the DELM, so the transfer is unvalidated.","rationale":"The reader and I identify the same weak point. It is load-bearing because the central new result is the scaling advantage: the experiments show no performance advantage, the simulations do, and the simulations' explanatory claim uses rank. If the rank-to-accuracy transfer fails, then the measured rank advantage in both experiment and simulation no longer supports the statement that enhanced expressivity leads to improved performance. The paper does provide direct simulated accuracy at n=3..6, which is genuine independent support, but it is one task with one detector count, and the claimed generality ('favorable scaling of expressivity with dimensionality') is carried by the rank analysis. Because the rank proxy is contradicted by the only experiment that measures both rank and accuracy, and because the rank threshold is arbitrary and noise-sensitive, the conditional verdict is appropriate rather than full acceptance. I do not see a reason to reject: the experimental reporting is honest, the model is standard, the simulation accuracy advantage is present, and the discrepancy is plausibly due to finite-sample/TM variability. The concrete test would settle whether the proxy survives.","tokens_in":9136,"tokens_out":6737,"duration_ms":75031,"concrete_test":"Run the simulation used for Fig. 4 over 100 random transmission matrices for n=3..6, m=2n, and for each TM compute both the linear-classifier accuracy on FashionMNIST and the feature-matrix rank at thresholds t=0.5, 0.7, 0.9, 0.99 and as a participation-ratio effective rank. Then test whether the IELM-vs-DELM rank ordering is stable across all t, and whether, across TMs within each photon-number class, rank predicts accuracy. If the ordering flips at high t, or if rank and accuracy are uncorrelated once the number of features is controlled, the expressivity explanation for the claimed quantum advantage fails; the paper would then need a direct accuracy-based demonstration at high photon number.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim is that indistinguishability yields a scalable QELM advantage, and the evidence offered for the mechanism is the increased rank of the feature matrix. That evidence only supports the claim if rank at threshold t=0.9 transfers to classification accuracy. The experiment provides a direct counterexample to this transfer: with 22 detectors the IELM has higher feature-matrix rank (Fig. 2c) while the DELM has higher accuracy, 93% vs 91% (Fig. 2a). The authors attribute this to dataset/TM variability, but the rank comparison itself has no error bar at 22 detectors ('there is only one subset of detectors, and thus we do not know the error bar'), and the threshold t=0.9 is not motivated. Moreover, the text concedes that shot noise inflates the measured rank toward the full random-matrix rank, so a higher rank can reflect noise rather than useful expressivity. The direct simulated accuracy gain at n>=3 (Fig. 4a) is real evidence for performance, but the abstract and conclusion assert the stronger causal statement that the gain arises from the rank/expressivity scaling; that causal link is exactly the unsupported rank-to-accuracy transfer, and the only direct experimental test of it goes the wrong way.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript reports an experimental photonic quantum extreme learning machine in which two photons are encoded with MNIST images, propagated through a multimode fiber, and detected on a SPAD array. Features are derived from intensities or from two-photon coincidences for distinguishable and indistinguishable photons. The experiments show that coincidence-based features outperform intensity-only features, while no significant difference is observed between distinguishable and indistinguishable photons at the experimental scale. Numerical simulations for up to five photons and for scaled detector numbers show an accuracy advantage for indistinguishable photons and a larger rank of the feature matrix, which the authors interpret as enhanced expressivity. The paper concludes that photon indistinguishability provides a scalable QELM advantage without relying on high-dimensional entanglement.","tokens_in":9281,"tokens_out":5489,"duration_ms":56273,"significance":"If the simulation results hold, the paper would provide evidence that photon indistinguishability can improve QELM accuracy at larger photon numbers without requiring entanglement. The experimental implementation is careful: Eq. (1) is the standard multimode bosonic coincidence formula, the comparison between DELM and IELM is reported with honest overlapping error bars, and the authors explicitly acknowledge the absence of a clear experimental advantage at the current scale. The rank-scaling analysis attempts to connect the simulated gain to a mechanistic cause, and the normalization of the rank against a Gaussian random matrix provides an external reference. However, as detailed below, the rank-to-accuracy proxy is not validated and is contradicted by the paper's own experimental data, so the causal claim needs substantial revision. The work is likely of interest to the quantum machine learning and photonic computing community.","major_comments":[{"comment":"At 22 detectors, the IELM has higher feature-matrix rank (Fig. 2c) while the DELM has higher classification accuracy, 93% vs 91% (Fig. 2a). The text states that 'The rank of the feature matrix serves as a practical proxy of system expressivity' and uses the rank to explain a performance advantage, but the only direct experimental test of this transfer goes in the wrong direction. Since the central claim attributes the simulated quantum advantage to increased rank and expressivity, the manuscript must either validate the rank-to-accuracy transfer in a setting where both quantities are measured, or substantially weaken the causal claim and present the rank scaling only as a heuristic that is not yet evidenced to predict classification performance.","section":"Experimental results with two photons and 22 detectors (Figs. 2a and 2c)"},{"comment":"The rank at threshold t=0.9 is sensitive to shot noise, as the text concedes: 'a high amount of shot noise in the system will add up in the measured feature matrix and the rank will converge to the rank of a random matrix'. The experimental rank at 22 detectors has no error bar ('there is only one subset of detectors, and thus we do not know the error bar'), and the threshold t=0.9 is not motivated. The higher experimental IELM rank could therefore be an artifact of noise or of the threshold choice rather than a sign of useful expressivity. Please provide a noise-model analysis of the rank for the experimental coincidence counts and a sensitivity study with respect to t.","section":"Variability of the system and Methods (rank calculation)"},{"comment":"The scaling claim that the distinguishable rank is linear in m while the indistinguishable rank is between linear and quadratic (and exponential for m=2n) needs clarification. For two photons, the dimension of the n-fold coincidence space is C(m,2) ~ m^2, so the distinguishable rank cannot exceed this dimension and a linear scaling is not obvious. The text does not state the number of random inputs p used to build the feature matrix in Fig. 4b,c; if p is comparable to m, the computed rank is truncated by p. Please specify p, the feature matrix dimensions, and the details of the Gaussian-matrix normalization, and verify that the reported scaling is not an artifact of matrix size or of the m=2n choice that simultaneously changes both photon number and detector number.","section":"Simulations when scaling the number of photons and modes (Fig. 4b,c)"}],"minor_comments":[{"comment":"The word 'densly' should be 'densely'.","section":"Abstract"},{"comment":"'Sterling's formula' should be 'Stirling's formula'.","section":"Simulations when scaling the number of photons and modes"},{"comment":"In the sentence 'The dimensionality of the n-fold coincidence space form detectors', 'form' should be 'for m detectors'.","section":"Simulations when scaling the number of photons and modes"},{"comment":"The coefficient α in Eq. (1) is called the indistinguishability, and Fig. 3a varies 'indistinguishability' from 0 to 1, but the relation between α, the measured HOM visibility, and the simulation parameter is not defined; please state explicitly how α enters the simulations and how it relates to experimental visibility.","section":"Equation (1) and Fig. 3a"},{"comment":"The caption of Fig. 4c should define the ordinate and state that the rank is normalized by the Gaussian random matrix rank, to match the description in the main text.","section":"Figure 4 caption"}],"recommendation":"major_revision","confidential_remarks":"The experimental contribution is solid and the authors are transparent about the lack of a clear experimental advantage. The main issue is the gap between the strong causal claim in the abstract and conclusion and the evidence supporting the rank-to-accuracy transfer. I believe this is addressable in revision by reframing the claim and adding supporting analysis, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new thing here is genuinely new: a coincidence-based QELM on a classical image task, built with a multimode fiber and a SPAD array, with distinguishable and indistinguishable photon pairs compared against intensity-only readout. The experimental realization is clean, the HOM visibility is solid, and the authors are honest about the ambiguous low-dimensional result: 93% for distinguishable, 91% for indistinguishable, overlapping error bars. The simulations also show a real, direct accuracy gain for indistinguishable photons once you go beyond two photons or scale the mode count, and the rank scaling slopes clearly differ between the two cases. I want to credit that direct evidence, because it is not nothing.\n\nThe soft spot is the load-bearing explanatory claim. The abstract and conclusion assert that the advantage arises from the increased rank of the feature matrix, but the only direct experimental test of that transfer goes the wrong way: at 22 detectors the IELM has higher rank, yet lower accuracy than the DELM. The authors wave it away as dataset or transmission-matrix variability, but it is a counterexample to the proxy. The rank threshold t=0.9 is arbitrary and unmotivated, there is no error bar on the 22-detector rank point, and the paper itself notes shot noise inflates the measured rank toward the full random-matrix rank—so a higher rank can simply mean more noise, not more useful expressivity. The simulated accuracy gain is real, but the claim that it arises from the rank scaling is not established.\n\nThe citation pattern is fine. Eq. (1) is standard, the model reference [14] is not self-cited, and the circularity burden is low. The missing code, data, and the dangling “supplementary Fig.??” references are sloppy but fixable.\n\nNet: the paper deserves a serious referee, but I would not accept it as-is. The authors need to either validate the rank-to-accuracy transfer on a larger-scale experiment or at least soften the causal story to match the evidence. Release the data and code, add a sensitivity analysis for the threshold, and fix the figure callouts. This is a credible experimental step, and the simulation results are worth engaging with, but the current framing oversells a proxy that the experiment does not yet support.","headline":"The paper's real contribution is a clean experimental demonstration of coincidence-based QELMs on a classical task, but its central claim about an indistinguishability-driven advantage rests on a rank-to-accuracy transfer that its own experiment contradicts.","tokens_in":10000,"tokens_out":1352,"would_cite":false,"duration_ms":15561,"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":"Photon indistinguishability—the property that lets identical photons interfere—is claimed to be a scalable resource that enlarges the feature space of photonic quantum extreme learning machines.","keywords":["photon indistinguishability","quantum extreme learning machine","multimode fiber","feature matrix rank","expressivity","Hong-Ou-Mandel interference","photon coincidences","random feature maps"],"falsifier":"Run the scaling simulation at $n=5$ photons with $m=10$ detectors across many random transmission matrices and compare actual test accuracy; if indistinguishable photons show no statistically significant accuracy gain over distinguishable photons despite a higher feature-matrix rank, the proposed link between expressivity and performance is refuted.","tokens_in":8867,"feed_emoji":"⚛️","tokens_out":7255,"duration_ms":67787,"temperature":0.7,"pith_summary":"This paper sets out to establish that photon indistinguishability—the property that lets identical photons interfere—can be a usable resource in quantum extreme learning machines (QELMs), a photonic variant of neural networks with a fixed random layer and only the readout trained. The authors build a two-photon QELM using a multimode fiber as the random layer and show experimentally that features built from photon coincidence measurements classify images better than features built from intensities alone. Their simulations then make the stronger prediction that as the photon number grows, indistinguishable photons outperform distinguishable photons by a widening margin, because the dimensionality of the feature space grows faster for indistinguishable photons. If the prediction holds, quantum interference alone—without high-dimensional entanglement—would give photonic learning machines a scalable expressivity advantage.","feed_headline":"Indistinguishable photons boost photonic machine learning at scale","feed_subtitle":"Coincidence-based quantum extreme learning machines gain expressivity as photon number grows, simulations show.","key_machinery":"The central object is the $n$-photon coincidence feature matrix, whose entries are the measured coincidences $C_{j_1\\cdots j_n}$ between output modes. For $n$ photons and $m$ detectors the ideal coincidence is $$C_{j_1\\cdots j_n}=\\frac{\\$\\alpha$}{n!}\\left|\\sum_{\\$\\sigma$\\in S(n)}E_{\\$\\sigma$(1)j_1}\\cdots E_{\\$\\sigma$(n)j_n}\\right|^2+\\frac{1-\\$\\alpha$}{n!}\\sum_{\\$\\sigma$\\in S(n)}\\left|E_{\\$\\sigma$(1)j_1}\\cdots E_{\\$\\sigma$(n)j_n}\\right|^2,$$ with $\\alpha=1$ for indistinguishable photons, $\\alpha=0$ for distinguishable photons, and $E_{ij}$ the output field of photon $i$ in mode $j$. The multimode fiber implements the fixed random projection; a single-photon avalanche diode (SPAD) array records intensities or two-photon coincidences; and the rank of the feature matrix—the number of singular values needed to reach 90% of the total squared energy—serves as the paper's proxy for the expressivity of the ELM. The scaling simulations use the coincidence Hilbert space whose dimension is the binomial coefficient $\\binom{m}{n}$.","core_discovery":"The paper's central claim is that indistinguishability can be harnessed as a computational resource in QELMs: with indistinguishable photons the random projections generated by a multimode fiber span a higher-dimensional feature space than with distinguishable photons, and this advantage increases with the number of photons. Experimentally, with two photons and up to 22 detectors, the coincidence-based ELMs reach 93% and 91% accuracy and clearly beat the 87% intensity-only ELM, but the distinguishable and indistinguishable variants are statistically comparable. The simulations extend the claim: fixing 16 detectors and going from one to five photons, the indistinguishable ELM rises from about 60% to 70% accuracy while the distinguishable one plateaus around 63%; the feature-matrix rank of indistinguishable photons scales between linear and quadratic in the number of detectors and, when $m=2n$, grows exponentially with photon number, versus a polynomial scaling for distinguishable photons. The paper concludes that the enhanced expressivity comes from quantum interference enriching the dimensionality of coincidence space, not from entanglement.","pith_inferences":["If the rank-to-accuracy transfer holds, the largest practical gains should appear on tasks that demand many independent features, such as fine-grained or texture classification, rather than the two-class MNIST task used here; the paper's own experiment may simply be too easy to expose the advantage.","The partial-indistinguishability optimum suggests that a variable-delay or tunable-coherence source could be treated as an adjustable knob and optimized per task; this is a control dimension not considered in the scaling simulations.","A direct comparison against classical random feature maps with the same number of output features would make the quantum advantage quantitative; the paper shows rank scaling but not a classical baseline at matched feature counts.","The experimental counterexample—higher rank but lower accuracy for the indistinguishable ELM—implies rank alone understates the role of generalization; a combined measure with the training-test accuracy gap would be a stricter expressivity proxy."],"forward_implications":["Coincidence-based QELMs are a practical upgrade over intensity-based optical ELMs: on the MNIST 0-1 task they improve accuracy from 87% to above 91% with the same optical hardware.","At currently accessible sizes (two photons, 22 detectors) indistinguishability is not yet a measurable advantage; the predicted advantage appears only as the number of photons and detectors grows.","The expressivity of indistinguishable-photon features scales faster than distinguishable-photon features—super-linear in detectors and exponential in photon number under $m=2n$—so photon number is a more efficient resource than adding detectors.","Partial indistinguishability ($\\alpha$ around 0.73) can outperform full indistinguishability on some tasks, making the coherence of the source a tunable hyperparameter rather than a maximized resource.","To exploit the predicted advantage at scale, future QELMs must avoid measuring the full exponential Hilbert space; the paper points to partial-measurement strategies as the needed direction."],"supporting_citations":[{"why":"It defines the extreme learning machine architecture with a fixed random hidden layer and linear readout that the photonic setup instantiates.","marker":"[5]"},{"why":"It supplies the optical random-projection method using scattering media that the multimode-fiber experiment builds on.","marker":"[8]"},{"why":"It introduces Hong-Ou-Mandel interference, the two-photon quantum effect that indistinguishability harnesses.","marker":"[12]"},{"why":"It provides the theoretical QELM analysis used to argue that the expressivity gain from interference persists even with partial Hilbert-space measurements.","marker":"[13]"},{"why":"It provides the coincidence Hilbert-space model and scaling setup used in the simulations as the number of photons and detectors grows.","marker":"[14]"},{"why":"It supplies the MNIST dataset used in the two-photon classification experiment.","marker":"[17]"},{"why":"It supplies the FashionMNIST dataset used for the multi-photon scaling and classification simulations.","marker":"[23]"},{"why":"It describes the SPAD-array and multimode-fiber platform used to record photon coincidences.","marker":"[26]"}],"fun_headline_variants":["Indistinguishable photons give quantum machine learning an edge","Scaling quantum ML: indistinguishability drives feature space growth","Photon indistinguishability enhances quantum extreme learning machines","Coincidence-based QELMs: indistinguishability improves photon scaling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on the assumption that the rank of the feature matrix—how many independent directions the random projections produce—faithfully tracks classification accuracy, even though in the two-photon experiment indistinguishable photons had higher rank while distinguishable photons had higher accuracy.","fun_headline_variants_meta":{"raw":{"variants":["Indistinguishable photons give quantum machine learning an edge","Scaling quantum ML: indistinguishability drives feature space growth","Photon indistinguishability enhances quantum extreme learning machines","Coincidence-based QELMs: indistinguishability improves photon scaling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00086,"raw_usage":{"total_tokens":3693,"prompt_tokens":870,"completion_tokens":2823,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":486,"completion_tokens_details":{"reasoning_tokens":2754}},"tokens_in":486,"tokens_out":2823,"duration_ms":19947,"temperature":1.0,"reasoning_tokens":2754,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:55:13.191932+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the scaling simulation at $n=5$ photons with $m=10$ detectors across many random transmission matrices and compare actual test accuracy; if indistinguishable photons show no statistically significant accuracy gain over distinguishable photons despite a higher feature-matrix rank, the proposed link between expressivity and performance is refuted.","supporting_citations":[{"cited_title":"Ex- treme learning machine: Theory and applications.Neurocom- puting, 70(1):489–501, December 2006","cited_arxiv_id":null,"evidence_quote":"It defines the extreme learning machine architecture with a fixed random hidden layer and linear readout that the photonic setup instantiates."},{"cited_title":"Saade, F","cited_arxiv_id":null,"evidence_quote":"It supplies the optical random-projection method using scattering media that the multimode-fiber experiment builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It introduces Hong-Ou-Mandel interference, the two-photon quantum effect that indistinguishability harnesses."},{"cited_title":"Lecun, L","cited_arxiv_id":null,"evidence_quote":"It supplies the MNIST dataset used in the two-photon classification experiment."},{"cited_title":"Large Reconfigurable Quan- tum Circuits with SPAD Arrays and Multimode Fibers, May","cited_arxiv_id":null,"evidence_quote":"It describes the SPAD-array and multimode-fiber platform used to record photon coincidences."}],"review_version":1}