{"id":"bf4dfa48-1cd6-4617-aec1-3bcac5998517","arxiv_id":"2507.12505","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A hybrid quantum-classical CNN for causality classification works only with a Pauli XYZ feature map, and deeper quantum ansatzes appear to act as implicit regularizers in single-run experiments.","lead":"This thesis adds a small quantum circuit to a classical convolutional neural network and tests nine different quantum feature encodings on a causality classification task. Only one encoding, Pauli XYZ, reached about 90% validation accuracy, while the others stayed near chance, showing that feature map choice can decide success or failure of a hybrid model.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No classical baseline or seeded repeats appear in §2.2–§3, so the claimed quantum benefit and ansatz regularization in Tables 3.1–3.5 cannot be attributed to the quantum component rather than to model capacity or chance.","rationale":"I agree with the reader's weakest assumption: the absence of a classical baseline is the critical gap. I would keep the REJECT verdict. The paper's own language in §2.2.6 and §2.3 frames the quantum layer as the intervention, but no intervention is compared against a no-quantum control. Moreover, the ansatz-depth experiment conflates depth with parameter count; because the classical feature extractor is fixed, any improvement from deeper ansatz could come from added trainable parameters. The single-run design compounds this: the stability ratio is derived from one trajectory, so sampling noise is not quantified. The proposed test is feasible because the code is public. If the matched-capacity classical baseline performs as well, the paper's central claim fails; if the hybrid is consistently better across seeds, the claim becomes credible. I found no internal inconsistency in the metrics themselves, but the evidential link from results to conclusion is missing.","tokens_in":41163,"tokens_out":3208,"duration_ms":38271,"concrete_test":"Using the released Zenodo/GitLab code, fix data splits, optimizer, and 500 epochs, and run at least five seeds for: (A) the reported HQCNN with three ansatz repetitions; (B) the same CNN and classifier with the quantum block removed (dimension-reduction output fed directly to the final linear layer); (C) a classical MLP inserted in place of the quantum block with the same number of trainable parameters as the three-repetition ansatz. If B or C matches or exceeds A's validation accuracy of 0.9111 and stability ratio of 0.8335, the quantum-specific component is not needed for the reported effect; conversely, if A consistently beats both matched-capacity baselines, the claim gains support.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the observed improvements—validation accuracy 0.8721→0.9111 and stability ratio 2.1791→0.8335 in Tables 3.1 and 3.5—are caused by the quantum circuit. The experimental design in §2.2.6 specifies only that each configuration was 'independently trained and evaluated'; it does not include a classical CNN without a quantum layer, nor any repetition of the same configuration with different seeds. The TwoLocal ansatz's repetitions add trainable parameters, so the deeper-ansatz benefit is exactly the kind of effect that a larger classical layer could produce. The epoch-level fluctuation metrics are computed from a single training trajectory, so the stability ratio has no error bar and cannot support a regularization claim. Without an ablation that holds capacity constant, the conclusion that 'quantum components can introduce meaningful transformations' and that ansatz repetition 'acts as regularization' is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript, formatted as a diploma thesis, investigates hybrid quantum-classical convolutional neural networks for a three-class causality-direction classification task on synthetic 8×8 heatmaps. It compares ansatz depths (one, two, and three repetitions of a TwoLocal circuit) and nine quantum feature maps, reporting validation accuracy, generalization gap, epoch-level fluctuation metrics, PCA silhouette scores, and Fisher discriminant ratios. The central claim is that increasing ansatz repetitions acts as a regularizer that improves generalization and that quantum components introduce meaningful transformations; only the Pauli XYZ feature map is reported to enable successful learning.","tokens_in":41347,"tokens_out":5687,"duration_ms":64114,"significance":"The question of whether small quantum layers can provide any benefit over classical feature extractors in hybrid models is timely and relevant to NISQ-era quantum machine learning. The manuscript's strengths include a fully specified architecture, a publicly available implementation on Zenodo/GitLab, and the use of several complementary diagnostics. However, the experimental design lacks a classical baseline, any repeated runs or seed variation, and capacity-matched ablations; as a result, the central enhancement and regularization claims are not supported by the current evidence. The PCA/Fisher analyses also contain internal inconsistencies that further weaken the conclusions.","major_comments":[{"comment":"No purely classical CNN baseline is trained, so the claimed benefit of the quantum layer is inferred only from comparisons among hybrid configurations. Because increasing ansatz repetitions from one to three also increases the number of trainable parameters, the observed validation accuracy improvement (0.8721 to 0.9111) and stability ratio reduction (2.1791 to 0.8335) could equally result from increased model capacity or from a favorable random initialization; an ablation holding capacity constant or a classical baseline is required to attribute these effects to the quantum circuit.","section":"§2.2.6, Tables 3.1–3.5"},{"comment":"Each configuration is trained exactly once, and the fluctuation metrics and stability ratio are computed from a single training trajectory. With no repeated seeds or variance estimates, the regularization interpretation rests on three uncorrelated single-run points and cannot be distinguished from random variation.","section":"§2.2.6, Tables 3.4–3.5"},{"comment":"The ranking of feature maps by silhouette score uses the 'Fitted values' (model predictions) rather than the true 'Training values' labels. High silhouette on the model's own predictions is tautological, and the true-label silhouette values in Table 3.9 are near zero for eight of nine maps; the table therefore does not measure actual class separability.","section":"§3.3, Tables 3.7–3.10"},{"comment":"The reported Fisher discriminant ratios are not symmetric under class permutation, contradicting the symmetric formula in Eq. (3.11). For example, for zz_feature_map_reps_1_linear, the '0 vs 1' entry is 16.9932 while the '1 vs 0' entry is 12.2353; this indicates a computational or definitional mismatch that invalidates the Fisher-based comparison.","section":"Table 3.11"}],"minor_comments":[{"comment":"The word 'achqcnn' appears to be a typo for 'the HQCNN model'.","section":"§2.5"},{"comment":"The text contains several typos, including 'expandion' for 'expansion', and the equation after 'ESA(θ) =' is not displayed consistently.","section":"§1.16"},{"comment":"The phrase 'The first attempt is visible in??' contains a broken cross-reference and should point to the specific figure.","section":"§3.3"},{"comment":"Sections 1.5, 1.14–1.16 (fermionic mappings, VQE, VHA, SA-OO-VQE) are not used in the experiments and could be trimmed or explicitly connected to the central study.","section":"Chapter 1"}],"recommendation":"reject","confidential_remarks":"The manuscript is a diploma thesis on a promising topic, but the empirical core is not yet at the level of a journal publication. The missing classical baseline and single-run design are the central obstacles; a substantial revision with new experiments would be needed. I recommend rejection because the current evidence does not support the stated conclusions, not because the underlying idea is uninteresting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Honest take: this is a master's thesis-level empirical scan of nine Qiskit feature maps and three TwoLocal ansatz depths on one causality classification dataset. The code is public, the tables are internally consistent, and the author is clearly transparent about what was run. That is real value: someone planning a similar hybrid pipeline can see in one place how sensitive these models are to feature map choice. The observation that only the Pauli XYZ feature map worked, and that the rest collapsed or failed, is a useful data point, even if it is just one dataset. The PCA and silhouette diagnostics are a nice touch; they show where separability does and does not emerge.\n\nThe soft spot is load-bearing. The thesis compares hybrid configurations against each other but never trains a classical CNN without the quantum layer. So the claim that 'quantum components can introduce meaningful transformations' has no external reference. Adding ansatz repetitions also adds trainable parameters; a larger classical head could plausibly produce the same stabilization. And with one run per configuration, the stability ratio is computed from a single trajectory—there is no seed variance. So the 0.8721 to 0.9111 validation accuracy and the 2.18 to 0.83 stability ratio are real numbers, but they do not support the regularization claim with any statistical grounding. The conclusion in the abstract and Chapter 3 overstates what the evidence can carry.\n\nThat said, I do not think this deserves hostile rejection. It is a well-organized empirical starting point. If the authors run a classical baseline, add seeded repeats, and maybe one more dataset, the ansatz-depth result could become a solid short conference paper. As it stands, the central claim is not established, but the work is honest and reproducible.\n\nRecommendation: send it to review if you want a careful look at hybrid QML empirical practice, but expect heavy revision. The missing baseline is fixable and the single-run issue is fixable. I would not desk-reject it outright; there is enough substance and transparency to warrant referee time.","headline":"A transparent hyperparameter scan of quantum feature maps and ansatz depths, but the missing classical baseline and single-run design leave the central quantum-regularization claim unsupported.","tokens_in":41851,"tokens_out":1887,"would_cite":false,"duration_ms":22196,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P68","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a small variational quantum circuit inside a classical CNN can improve generalization and stability, and that the feature map choice is decisive: of nine tested maps, only the Pauli XYZ map trains well.","keywords":["hybrid quantum-classical neural networks","variational quantum circuits","quantum feature mapping","quantum machine learning","convolutional neural networks","ansatz depth","regularization","causality classification"],"falsifier":"Train the same CNN once with the quantum block and once with a classical linear layer of the same parameter count in its place, keeping data, optimizer, and epochs identical; if the classical swap reproduces the three-repetition results (validation accuracy near 91.1 percent and a stability ratio near 0.83) across several random seeds, the attribution of the regularization effect to the quantum circuit is not supported.","tokens_in":40967,"feed_emoji":"⚛️","tokens_out":13691,"duration_ms":151973,"temperature":0.7,"pith_summary":"The paper tries to establish that a small variational quantum layer can do real work inside an otherwise classical convolutional network: on a three-class causality task, a quantum circuit with only a few qubits placed between the feature extractor and the classifier changes not just final accuracy but the whole training dynamics. Repeating the quantum ansatz from one to three times raises validation accuracy from 87.2 to 91.1 percent, shrinks the generalization gap, and lowers the validation-to-training fluctuation ratio from 2.18 to 0.83, which the author interprets as the deeper circuit acting like an implicit regularizer. The second claim is that the quantum feature map is the dominant design choice: of nine maps tested, only the Pauli XYZ rotation map produced a well-trained model, while others collapsed the output or failed to separate classes. The reason to care is practical: if a few-qubit quantum layer can steady a CNN's training, hybrid models become a cheap add-on to existing pipelines rather than a wholesale replacement.","feed_headline":"Quantum layer depth lifts CNN validation accuracy to 91 percent","feed_subtitle":"Ansatz repetition from one to three also cuts the validation-fluctuation ratio from 2.18 to 0.83.","key_machinery":"The central object is the Hybrid Quantum-Convolutional Neural Network (HQCNN): a classical convolutional feature extractor (three blocks of convolution, ReLU, max-pooling, and dropout) whose flattened output is projected onto a small number of qubits, fed through a quantum neural network built from a data-encoding feature map followed by a hardware-efficient TwoLocal ansatz, a parameterized circuit of single-qubit rotations interleaved with entangling gates, and finally passed to a classical linear classifier. The two independent knobs that carry the argument are the number of ansatz repetitions, which deepens the parameterized circuit and smooths the learning curve, and the choice of feature map, which decides how classical inputs are embedded into Hilbert space (the tested variants use Z rotations, ZZ interactions, and multi-axis Pauli rotations with different entanglement patterns). The quantum layer is wired into the classical training loop through a differentiable connector, so gradients flow back through the circuit and the whole model trains end-to-end with ordinary gradient descent. The supporting diagnostics are PCA applied at three stages of the network, silhouette scores for cluster quality, the Fisher Discriminant Ratio for class separability, and the stability ratio, the mean absolute fluctuation of validation accuracy divided by that of training accuracy, which falls from 2.18 to 0.83 as ansatz repetitions grow from one to three.","core_discovery":"On the paper's own terms, the central claim is that quantum components can introduce meaningful transformations even with a limited number of qubits, and that the depth of the variational ansatz shapes the model's generalization as much as its accuracy. With three ansatz repetitions the hybrid model reaches 90.1 percent training and 91.1 percent validation accuracy, the smallest final generalization gap (0.0107), an overfitting drop of only 0.68 percent, and a stability ratio of 0.83, meaning validation accuracy fluctuates less than training accuracy; the author concludes that repeating the ansatz acts as a form of regularization. The feature-map study carries equal weight: only the Pauli XYZ map with one repetition learns successfully (90.1 percent validation accuracy), and the paper shows through PCA and silhouette analysis that the quantum layer of this model compresses the data into a single well-separated component, while failed maps either collapse the class structure or leave it unrecoverable. The overall conclusion is that the design of the quantum part, the feature map and the ansatz together, determines learning dynamics, stability, and final performance in hybrid models, and that increased complexity must be introduced deliberately.","pith_inferences":["A testable consequence the paper leaves implicit is that a classical layer of matched parameter count might reproduce the observed regularization, since no purely classical baseline was trained; until that control exists, the effect cannot be uniquely attributed to quantum processing.","The stability-ratio pattern resembles what stochastic regularizers such as dropout produce, which suggests a direct control experiment: replace the quantum block with a classical stochastic layer of comparable noise and compare the fluctuation curves.","The paper's own warning that the Fisher Discriminant Ratio can look good for ill-trained models implies that separability metrics should be checked against true labels rather than fitted values; a screening protocol that tests label-agreement of quantum-layer outputs before training would be a cheap extension.","All reported comparisons appear to come from single runs per configuration, so multi-seed replication with statistical testing is the natural next experiment before the 87-to-91 percent improvement is treated as firm."],"forward_implications":["Practitioners should treat ansatz depth as a tunable regularizer in hybrid classifiers: three repetitions improve generalization metrics even though raw accuracy gains are modest, and the stability ratio drops below one.","Feature-map choice can dominate circuit size: models with Z-only or heavily entangled maps fail to separate classes at all, so map selection deserves the same care as architecture selection.","Early plateaus in deeper quantum circuits are part of the learning process, not a defect: the three-repetition model had a negative early slope yet was the first to exceed 90 percent validation accuracy, at epoch 168.","Blindly adding feature-map complexity can cause dimensional collapse, so complexity should be introduced with intent and monitored.","Evaluating hybrid models by final accuracy alone understates the trade-offs: fluctuation, generalization gap, and stability ratio are needed to see what deeper quantum layers contribute."],"supporting_citations":[{"why":"Supplies the causality-classification task and the classical deep-network approach that this thesis extends by inserting quantum layers.","marker":"[126]"},{"why":"Grounds the variational quantum circuit training loop (parameterized ansatz, measurement, classical optimizer) that the hybrid layer is built on.","marker":"[101]"},{"why":"Cited for the claim that quantum-layer placement shapes hybrid model behavior, motivating the architecture comparisons.","marker":"[99]"},{"why":"Frames hybrid quantum-classical models as the viable near-term (NISQ) approach that the experiments operate within.","marker":"[95]"},{"why":"The author's prior variational-eigensolver work that motivates the progression into hybrid quantum-classical machine learning.","marker":"[125]"},{"why":"Grounds the premise that quantum parallelism is the resource a quantum layer contributes to the classical network.","marker":"[100]"},{"why":"Provides the angle-encoding design basis for the feature maps tested in the experiments.","marker":"[96]"},{"why":"Supplies the PCA method used to diagnose data separability at each network stage.","marker":"[127]"}],"fun_headline_variants":["Quantum layers boost CNN accuracy to 91%","Ansatz repetition regularizes hybrid quantum CNN","Feature map choice makes or breaks quantum CNN","Hybrid quantum CNN: less fluctuation, more stability","Quantum layers in CNNs: small qubit count, big impact"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Because the experiments never train a purely classical network without the quantum layer, the claim that the observed accuracy and stability gains come from the quantum circuit itself rests on comparing hybrid configurations against each other rather than against a classical baseline.","fun_headline_variants_meta":{"raw":{"variants":["Quantum layers boost CNN accuracy to 91%","Ansatz repetition regularizes hybrid quantum CNN","Feature map choice makes or breaks quantum CNN","Hybrid quantum CNN: less fluctuation, more stability","Quantum layers in CNNs: small qubit count, big impact"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000614,"raw_usage":{"total_tokens":2828,"prompt_tokens":894,"completion_tokens":1934,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":510,"completion_tokens_details":{"reasoning_tokens":1860}},"tokens_in":510,"tokens_out":1934,"duration_ms":15157,"temperature":1.0,"reasoning_tokens":1860,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:48:12.397680+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same CNN once with the quantum block and once with a classical linear layer of the same parameter count in its place, keeping data, optimizer, and epochs identical; if the classical swap reproduces the three-repetition results (validation accuracy near 91.1 percent and a stability ratio near 0.83) across several random seeds, the attribution of the regularization effect to the quantum circuit is not supported.","supporting_citations":[{"cited_title":"Causal inference using deep neural networks","cited_arxiv_id":"2011.12508","evidence_quote":"Supplies the causality-classification task and the classical deep-network approach that this thesis extends by inserting quantum layers."},{"cited_title":"The variational quantum eigensolver: a review of methods and best practices.Physics Reports","cited_arxiv_id":null,"evidence_quote":"Grounds the variational quantum circuit training loop (parameterized ansatz, measurement, classical optimizer) that the hybrid layer is built on."},{"cited_title":"Challenges and opportunities in quantum machine learning.Nature computational science","cited_arxiv_id":null,"evidence_quote":"Cited for the claim that quantum-layer placement shapes hybrid model behavior, motivating the architecture comparisons."},{"cited_title":"A survey of NISQ era hybrid quantum-classical machine learning re- search","cited_arxiv_id":null,"evidence_quote":"Frames hybrid quantum-classical models as the viable near-term (NISQ) approach that the experiments operate within."},{"cited_title":"Transformation-free generation of a quasi-diabatic representation from the state- average orbital-optimized variational quantum eigensolver.arXiv preprint arXiv:2502.18194","cited_arxiv_id":null,"evidence_quote":"The author's prior variational-eigensolver work that motivates the progression into hybrid quantum-classical machine learning."},{"cited_title":"What is quantum parallelism, anyhow? In:ISC High Performance 2024 Research Paper Proceedings (39th International Conference)","cited_arxiv_id":null,"evidence_quote":"Grounds the premise that quantum parallelism is the resource a quantum layer contributes to the classical network."},{"cited_title":"Quantum angle encoding with learnable rotation applied to quantum–classical convolutional neural networks.Applied Soft Computing","cited_arxiv_id":null,"evidence_quote":"Provides the angle-encoding design basis for the feature maps tested in the experiments."},{"cited_title":"Principal component analysis.Wiley interdisciplinary reviews: computational statistics","cited_arxiv_id":null,"evidence_quote":"Supplies the PCA method used to diagnose data separability at each network stage."}],"review_version":1}