{"id":"402ac4ec-705b-4ba0-bdbd-966bb688e4c3","arxiv_id":"2502.08968","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"Quanvolutional neural networks outperformed classical CNNs on dysphonia classification from small Mel-spectrogram datasets, though the experimental design is flawed.","lead":"Researchers compared quantum-inspired 'quanvolutional' neural networks with standard CNNs for detecting dysphonia from voice recordings. The quantum models achieved higher average accuracy on a small dataset of vowel sounds, but methodological flaws make the comparison unreliable.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test-set contamination via early stopping on the 61 held-out samples is load-bearing; the QNN-versus-CNN comparison is not a clean out-of-sample test.","rationale":"The reader's identified weakest assumption is the same as the most load-bearing concern here. For the abstract claim to be true, the QNN-versus-CNN comparison must be a valid out-of-sample comparison. That requires model selection (early stopping) to use information that does not include labels or losses from the 61 test samples. The paper's own description contradicts this requirement: Section III-B explicitly says early stopping uses the testing loss, and the experimental section repeatedly refers to testing accuracy and testing loss on the fixed 61-sample test set. Because the checkpoint is chosen by monitoring the exact samples used for final evaluation, the reported means and standard deviations cannot be interpreted as independent estimates of generalization. No formal verification, code release, or significance testing is provided to compensate. The random quantum circuit is not specified precisely enough to reproduce or to rule out chance, but that is secondary; the test-set early stopping alone makes the central claim unreliable. Since this is the same decisive weakness the reader used to reject the paper, the disposition remains REJECT; no verdict adjustment is needed.","tokens_in":7672,"tokens_out":5991,"duration_ms":57397,"concrete_test":"Run the authors' pipeline with early stopping based on a validation set disjoint from the 61 test samples—e.g., hold out 20% of the 243 training samples as validation for each training size, repeat the same 10 runs, and evaluate on the untouched 61 samples. Compare mean test accuracy of QNN1/CNN1 and QNN2/CNN2. If the claimed QNN advantage (76-85% vs 73-75% at 60-160 samples) persists with non-overlapping confidence intervals, the concern is resolved; if the gap shrinks or reverses, the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is unsupported because the evaluation procedure appears to use the test set for model selection. Section III-A describes only a 243/61 train/test split, with no validation set. Section III-B states: 'Early Stopping applied to halt training if the testing loss does not decrease after 15 consecutive epochs,' and Figures 5-6 plot 'Testing accuracy' and 'Testing loss' versus epoch. Thus the same 61 test samples are used to select the stopping point and to compute the reported accuracy. Under this protocol, the reported accuracy is a selection statistic, not a fair generalization estimate, and any QNN advantage—including the stability claim based on smaller standard deviations—could be an artifact of checkpoints chosen on the test set. The unspecified random quantum circuit in Section II-B.1.b is also a reproducibility blocker, but the test-loss early stopping is decisive: it invalidates the comparison even if the circuit were fully specified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes hybrid quantum-classical Quanvolutional Neural Networks (QNNs) for dysphonia classification from Mel spectrograms and compares them with classical CNNs on a small speech dataset (PVQD, 304 samples split into 243 training and 61 test images). Two QNN and two CNN architectures are evaluated across ten training-set sizes (60 to 240 samples), with the reported claim that QNNs consistently achieve higher accuracy and lower variance than CNNs. The quanvolutional layer uses 2x2 patches, angle encoding, a random quantum circuit, and Pauli-Z expectation decoding, simulated in PennyLane. The paper reports mean test accuracy and standard deviation over ten experiments, plus epoch-wise accuracy and loss curves for two training sizes.","tokens_in":7814,"tokens_out":4192,"duration_ms":42521,"significance":"If the reported advantage were methodologically sound, the paper would provide a useful early data point on whether quanvolutional feature extractors help in small medical audio datasets, which is a plausible and interesting hypothesis. The study has some strengths: it addresses a clinically relevant problem, uses a public dataset, varies the training-set size systematically, and keeps the compared architectures nearly identical except for the first feature-extraction layer. However, the manuscript does not ship code, does not specify the random quantum circuit, and, most importantly, appears to use the held-out test loss for early stopping. Because the central comparison is therefore not a clean out-of-sample evaluation, the quantitative conclusions about QNN superiority are not supported as presented. The contribution is better viewed as a preliminary benchmark idea that would need a properly designed validation protocol, a fully specified quantum circuit, and statistical testing to be credible.","major_comments":[{"comment":"Section III-A describes only a 243/61 train/test split with no validation set, and Section III-B states that early stopping halts training if the 'testing loss' does not decrease after 15 consecutive epochs, with Figures 5-6 plotting 'Testing accuracy' and 'Testing loss' versus epoch. If this loss is computed on the same 61 held-out samples used to report final accuracy, then model selection is performed on the test set and the reported accuracies are selection statistics rather than unbiased generalization estimates. The QNN-versus-CNN comparison in Figure 4 is therefore not a clean test of the abstract's claim. If the authors instead mean a validation loss, the manuscript does not describe where that validation set comes from; either way, the current protocol does not support the stated conclusion.","section":"III-A / III-B"},{"comment":"The random quantum circuit U is described only as choosing randomly among 1-qubit gates (Rx, Ry, Rz, T, H) and 2-qubit gates (CNOT, SWAP, CZ) with random angles. The number of circuit layers, the distribution over gates, the parameter ranges, the entanglement structure, and the random seed are not specified. Consequently, the QNN implementation is not reproducible. Moreover, if the quanvolutional layer is fixed before the classical layers are trained, the comparison may be between a random feature extractor and a trained CNN, which would not support the paper's attribution of the advantage to quantum feature extraction. Please specify the circuit construction and, ideally, report results averaged over multiple random circuit instances.","section":"II-B.1.b"},{"comment":"The paper reports mean accuracy and standard deviations over '10 folds' but provides no statistical significance tests, confidence intervals, or paired comparisons. With only 61 test samples, a difference of a few percentage points between QNNs and CNNs (Figure 4) can easily arise from sampling noise; the claim that QNNs are 'consistently' superior is not quantified. Please report per-experiment test accuracies, exact p-values or bootstrap confidence intervals, and the number of independent test samples contributing to each point.","section":"IV"},{"comment":"The experimental protocol is underspecified: the relationship between the '10 experiments' with increasing training sizes (60-240), the '10 training steps', and the '10 folds' mentioned in Section III-B is not explained. For example, are the same 61 test samples reused across all ten experiments, and are the '10 folds' cross-validation folds within the training portion or repeated random draws? This ambiguity makes it impossible to determine the effective sample size and the independence of the ten accuracy values whose mean and standard deviation are plotted in Figure 4.","section":"III-A / III-B"}],"minor_comments":[{"comment":"The word 'Comparision' should be 'Comparison'.","section":"Figure 5 caption"},{"comment":"The phrase 'height-by-weight-by-channel' should read 'height-by-width-by-channel' for the input dimension 40-by-100-by-1.","section":"II-A"},{"comment":"It should be stated explicitly whether the 61 test samples are disjoint from the randomly selected training subsets in every experiment; as written, a reader cannot tell whether the same patient or recording can appear in both the training subset and the test set across different runs.","section":"III-A"},{"comment":"The mention of 'k-fold Cross-Validation' is not accompanied by a description of how folds are formed; this is connected to major comment 4 and should be clarified.","section":"III-B"},{"comment":"No code, data split details, or random seeds are provided, which prevents exact replication of the experiments even if the quantum circuit were fully specified.","section":"General"}],"recommendation":"reject","confidential_remarks":"The paper addresses a timely application, but the evaluation protocol is not sound: the apparent use of test loss for early stopping invalidates the central comparison, and the unspecified random quantum circuit makes the QNN under test not a well-defined model. In my view, the changes needed (a proper held-out validation set, a fully specified and reproducible quantum circuit, and statistical testing) would require rerunning all experiments, which goes beyond a normal revision. I therefore recommend reject rather than major revision, although I would not discourage the authors from resubmitting a corrected version with a sound evaluation protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, it is a legitimate first application of quanvolutional neural networks to dysphonia detection, with a clean apples-to-apples architecture comparison—only the first convolution layer is swapped. Second, the central empirical claim is not supported because the early stopping criterion uses the test loss, so the 61 held-out samples are doing double duty as model selection and final evaluation.\n\nWhat the paper does well: the experimental design of varying training set size from 60 to 240 and reporting mean and standard deviation over repeated runs is sensible for a small-data question. The QNN method comes from Henderson et al. 2020, and the authors cite it properly, along with Yang et al. 2021 for speech. They are honest about their setup and provide version numbers for PennyLane and TensorFlow. The transfer to a voice disorder dataset is a reasonable extension, and the paper is clearly written.\n\nThe soft spots are load-bearing. Section III-B says early stopping halts training if the testing loss does not decrease for 15 epochs, and Figures 5-6 plot testing loss and testing accuracy against epochs. With no validation set, the model is being selected on the same 61 test samples used to report accuracy. That makes the reported accuracies selection statistics, not clean out-of-sample evaluations. The QNN advantage, especially the stability claim from smaller standard deviations, could be an artifact of picking checkpoints on the test set. Second, the random quantum circuit is described only as random gates with random angles; no seed, distribution, or concrete circuit is given, so the experiment is not reproducible. Third, there are no significance tests across the runs. The first issue alone invalidates the headline result.\n\nAll that said, this is a fixable paper, not a hopeless one. A revised version with a proper validation split for early stopping, a specified circuit, and significance tests could be a useful data point for the small-data quantum ML literature. As is, I would not cite the accuracy comparison, but I would send it to review with a request for major revision rather than desk-reject: the question is worth asking and the flaw is methodological rather than fatal to the idea.","headline":"A legitimate first application of QNNs to dysphonia, but early stopping on the test set leaks evaluation information and voids the headline accuracy comparison.","tokens_in":8356,"tokens_out":2536,"would_cite":false,"duration_ms":24408,"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":"Quantum feature extraction beats plain CNNs for dysphonia detection when voice data is scarce.","keywords":["quantum machine learning","quanvolutional neural network","dysphonia detection","Mel spectrogram","small speech dataset","hybrid quantum-classical model","voice disorder classification","early stopping"],"falsifier":"Re-run the ten training-set-size experiments with early stopping based on a separate validation set rather than the test set, then evaluate both matched architectures on the same 61 held-out samples; if the QNN no longer beats the CNN in mean accuracy or standard deviation, the central claim is refuted. A second check is to repeat each experiment under many random seeds and random quantum circuits and ask whether the advantage survives the resulting variance.","tokens_in":7465,"feed_emoji":"⚛️","tokens_out":8894,"duration_ms":82519,"temperature":0.7,"pith_summary":"The paper aims to show that a hybrid quantum-classical neural network can detect dysphonia from voice recordings more accurately and more stably than a classical CNN when labeled speech data is scarce. It builds paired architectures that differ only in the first feature-extraction layer, comparing two Quanvolutional Neural Networks (QNNs) against two otherwise identical CNNs on Mel spectrograms of sustained /a/ vowels. Ten experiments scale the training set from 60 to 240 samples while testing on the same 61 held-out files. In the paper's reported results, the QNN versions consistently show higher mean accuracy and smaller standard deviation, with the deeper QNN2 outperforming CNN2 at every training-set size. If the comparison is sound, it would give clinicians and researchers a concrete reason to use quantum feature extraction in small medical audio datasets.","feed_headline":"Quantum layers beat CNNs on scarce voice data","feed_subtitle":"Hybrid quantum-classical models give higher accuracy and stability for dysphonia detection on 60-240 samples.","key_machinery":"The carrying object is the quanvolutional layer, a drop-in replacement for a classical convolutional layer in which a small image patch is encoded into a quantum circuit rather than multiplied by a learned kernel. Here the layer uses $2\\times2$ patches, angle encoding via $R_y$ rotations with angles proportional to pixel intensities, a random entangling circuit, and Pauli-Z decoding into four feature channels. Because the paired CNN and QNN architectures differ only in this first layer, any systematic difference in accuracy and variance is attributed to the quantum feature extraction this layer performs.","core_discovery":"The central claim is that replacing the first convolutional layer with a $2\\times2$ quanvolutional layer improves dysphonia classification on small data. Each patch of the Mel spectrogram is angle-encoded onto four qubits, processed by a random quantum circuit whose two-qubit gates create entanglement, and read out as Pauli-Z expectation values to form four output channels; the rest of the network is classical. Across ten training-set sizes, QNN1 reaches 76%–85% mean accuracy for 60–160 training samples while CNN1 reaches 73%–75%, and QNN2 maintains higher accuracy with tighter standard deviations than CNN2 at every size. The authors interpret the result as evidence that quantum correlations extract features that are useful precisely when classical CNNs are data-hungry.","pith_inferences":["If early stopping was applied using the same 61 samples later used for final accuracy, the reported margins could partly come from test-set-aware model selection; repeating the runs with a separate validation split is the natural check.","The random quantum circuit is generated once per model, so the results may depend on that draw; rerunning across many random circuits and reporting the spread would show whether the advantage is stable.","The quanvolutional layer is simulated classically, so the experiment does not yet measure the effect of hardware noise on a real quantum processor; a NISQ implementation could erase or enlarge the gap.","The same first-layer swap could be tested on other scarce medical audio tasks, such as detecting Parkinsonian voice or COVID-19 cough, to see whether the quantum advantage generalizes beyond dysphonia."],"forward_implications":["Dysphonia can be classified from a single sustained /a/ vowel with roughly 76%–87% mean test accuracy using only 60–240 training spectrograms, which is a realistic data budget for a rare voice disorder.","Adding classical convolutional and pooling layers after the quanvolutional layer raises accuracy, so quantum feature extraction composes with ordinary deep-learning building blocks.","The reported advantage is largest at the smallest training sizes, suggesting the quantum layer is most useful exactly when data is hardest to obtain.","Because QNNs reach high test accuracy in fewer epochs, they also reduce the number of training iterations needed in simulation.","The same paired architecture could serve as a template for testing quantum feature extraction on other small medical audio datasets."],"supporting_citations":[{"why":"Introduces the quanvolutional layer and the hybrid architecture (quanvolution plus pooling and fully connected layers) that the paper adapts.","marker":"[20]"},{"why":"Previous use of $2\\times2$ quantum convolution layers on Mel spectrograms for speech recognition, the direct technical predecessor.","marker":"[19]"},{"why":"Provides the recent quanvolutional feature exploration and the simple QNN topology with fully connected layers used in the first scenario.","marker":"[24]"},{"why":"Supplies the public voice-quality database from which the sustained /a/ vowel training and test samples are drawn.","marker":"[25]"},{"why":"Provides the quantum machine learning simulation library used to implement and train the quanvolutional layers.","marker":"[26]"},{"why":"Supplies the early-stopping method used in the training protocol that determines when each model halts.","marker":"[27]"}],"fun_headline_variants":["Quantum-classical nets edge out CNNs on tiny voice datasets","Hybrid quantum model wins on small dysphonia data","QNNs trump CNNs when speech data is scarce","Quantum layer boosts dysphonia detection with few samples","Small data? Quantum hybrid beats classical CNN"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison stands only if the 'testing loss' used for early stopping is computed on a validation split that is separate from the 61 held-out samples used to report final accuracy, so model selection does not contaminate the reported QNN-versus-CNN comparison.","fun_headline_variants_meta":{"raw":{"variants":["Quantum-classical nets edge out CNNs on tiny voice datasets","Hybrid quantum model wins on small dysphonia data","QNNs trump CNNs when speech data is scarce","Quantum layer boosts dysphonia detection with few samples","Small data? Quantum hybrid beats classical CNN"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000189,"raw_usage":{"total_tokens":1292,"prompt_tokens":860,"completion_tokens":432,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":476,"completion_tokens_details":{"reasoning_tokens":354}},"tokens_in":476,"tokens_out":432,"duration_ms":4154,"temperature":1.0,"reasoning_tokens":354,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T23:02:51.584045+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the ten training-set-size experiments with early stopping based on a separate validation set rather than the test set, then evaluate both matched architectures on the same 61 held-out samples; if the QNN no longer beats the CNN in mean accuracy or standard deviation, the central claim is refuted. A second check is to repeat each experiment under many random seeds and random quantum circuits and ask whether the advantage survives the resulting variance.","supporting_citations":[{"cited_title":"Quanvolutional neural networks: powering image recognition with quantum circuits,","cited_arxiv_id":null,"evidence_quote":"Introduces the quanvolutional layer and the hybrid architecture (quanvolution plus pooling and fully connected layers) that the paper adapts."},{"cited_title":"Decentralizing feature extraction with quantum convolutional neural network for automatic speech recognition,","cited_arxiv_id":null,"evidence_quote":"Previous use of $2\\times2$ quantum convolution layers on Mel spectrograms for speech recognition, the direct technical predecessor."},{"cited_title":"Exploring the features of quanvolutional neural networks for improved image classification,","cited_arxiv_id":null,"evidence_quote":"Provides the recent quanvolutional feature exploration and the simple QNN topology with fully connected layers used in the first scenario."},{"cited_title":"Perceptual voice qualities database (pvqd),","cited_arxiv_id":null,"evidence_quote":"Supplies the public voice-quality database from which the sustained /a/ vowel training and test samples are drawn."}],"review_version":1}