{"id":"e037545b-3e5f-4754-9048-68fddbdb511f","arxiv_id":"2412.01241","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"Quantum pointwise convolution, a variational quantum circuit used as a 1x1 convolution kernel, is claimed to beat classical pointwise CNNs, but the evidence is confounded and unreproducible.","lead":"The paper proposes replacing the 1x1 (pointwise) convolution layer in a CNN with a parameterized quantum circuit, testing it on FashionMNIST and CIFAR10. It reports higher accuracy for the quantum version, but the comparison uses unmatched baselines, lacks error bars, and provides no code.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The empirical comparison is uncontrolled: the classical baseline has BatchNorm+ReLU after every layer while the quantum model does not, so the claimed quantum advantage may be an artifact of architectural mismatch and is never backed by reported parameter counts.","rationale":"The reader's weakest assumption identifies the same load-bearing issue: the classical and quantum baselines are not architecturally matched, since BatchNorm and ReLU appear only in the classical model. This matters because the paper's entire case for quantum pointwise convolution is the reported accuracy difference. In good faith, the rest of the paper is mostly a description of the proposed layer and a demonstration; the novelty lies in applying amplitude-encoded strongly entangling circuits as a pointwise-convolution replacement. For that application to be validated, the comparison must be controlled. It is not: no seeds or error bars are shown, no code is released, and no parameter counts are given despite the conclusion's explicit claim of 'significantly fewer parameters.' These are not merely presentation issues; they directly undercut the empirical claim. The concrete test I propose would isolate the source of the accuracy gap by adding/removing BatchNorm and ReLU symmetrically and by measuring seed variance and parameter counts. If the quantum advantage disappears in that controlled comparison, then the central claim collapses; if it persists, the paper's conclusion would be on much firmer ground. The verdict of REJECT is consistent with this analysis, so I recommend no change to the reader's verdict. I am not raising an objection to the existence of quantum pointwise convolution as a concept, nor to the theoretical possibility of quantum advantage, only to the sufficiency of the evidence presented.","tokens_in":9374,"tokens_out":2242,"duration_ms":21220,"concrete_test":"Reproduce the Fig. 3 architectures in PennyLane/PyTorch and train four matched conditions on FashionMNIST and CIFAR10 with identical optimizer, learning-rate schedule, batch size, epochs, and at least 5 random seeds: (1) classical as in Fig. 3b; (2) classical Fig. 3b with BatchNorm/ReLU removed to match the quantum model's normalization; (3) quantum as in Fig. 3a; (4) quantum Fig. 3a with BatchNorm (and ReLU where classically used) inserted after each layer. Report mean +/- std test accuracy and total parameter counts for every condition. If the quantum model still outperforms all classical controls by a margin larger than the seed noise and uses fewer parameters, the central claim survives; otherwise the advantage is explained by the architectural mismatch.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical: Section V's conclusion states that substituting classical convolution with quantum circuits yields equivalent or better performance with significantly fewer parameters. All evidence for this is training-curve comparisons (Figs. 4-8) with no error bars, no seeds, no code, and no quantified parameter counts. The load-bearing premise is that the two architectures in Fig. 3 are matched counterparts. They are not: the classical model (Fig. 3b) inserts BatchNorm and ReLU after every 1x1 and 3x3 layer, while the quantum model (Fig. 3a) has no BatchNorm and only one ReLU after the 3x3 layer. BatchNorm changes the optimization landscape and generalization behavior, so the reported gaps (Fig. 7: roughly 90% quantum vs. 80% classical on CIFAR10) could be caused entirely by this normalization/activation mismatch rather than by quantum pointwise convolution. The 'significantly fewer parameters' clause is likewise unsupported because neither model's parameter count is ever reported. The channel-count arithmetic is also internally inconsistent (Section III.E states Cout = ncircuits x nqubits, yet the Fig. 3 description and figures suggest channel counts that are not reconciled with n_qubits=6). Because the empirical comparison is the only evidence for the central claim, this uncontrolled baseline makes the claim unsubstantiated as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a 'quantum pointwise convolution' layer that replaces a classical 1x1 convolution: for each spatial position, the channel vector is amplitude-encoded into a few qubits, processed by a strongly entangling parameterized circuit, and measured in the Pauli-Z basis, with several circuits concatenated to form output channels. The authors embed this layer in a small CNN (Fig. 3a) and compare it with a classical counterpart (Fig. 3b) on FashionMNIST and CIFAR10. They report faster loss decrease and higher accuracy for the quantum model, and conclude in Section V that substituting classical convolution with quantum circuits yields equivalent or better performance with significantly fewer parameters. A second experiment varies the number of entangling layers.","tokens_in":9644,"tokens_out":5240,"duration_ms":44890,"significance":"The idea of using a parameterized quantum circuit as a channel-mixing 1x1 convolution is coherent and potentially useful: it maps naturally onto existing CNN architectures, and the design choices (amplitude encoding, weight sharing, one circuit per kernel, multiple measured qubits per circuit) are reasonable. If the empirical comparison were properly controlled, a demonstration that a quantum pointwise layer matches or beats a classical pointwise layer at equal or lower parameter count would be of interest to the quantum machine learning community. However, the current paper does not deliver that demonstration: the baseline is architecturally mismatched, no error bars or seeds are reported, the reported classical accuracy is internally inconsistent, and the claimed parameter reduction is never quantified. The contribution is therefore conditional on a substantially strengthened experimental section.","major_comments":[{"comment":"The comparison in Fig. 3 is not matched. The classical model in Fig. 3(b) has BatchNorm and ReLU after every convolution, whereas the quantum model in Fig. 3(a) has no BatchNorm and only one ReLU after the 3x3 layer. Since BatchNorm changes the optimization landscape and generalization, the accuracy deltas in Figs. 4-7 are not attributable to the quantum pointwise layer. The authors should compare architectures that differ only in the pointwise layer (both with or both without BN/ReLU, with the same placement), and should state explicitly that the comparison is controlled in this sense.","section":"Section IV, Fig. 3"},{"comment":"The reported classical accuracy is inconsistent: the text for Fig. 7 states that the classical model stabilizes around 80%, while the text for Fig. 8 states that the classical model (same dataset, same architecture) reaches a peak of only about 70%. This discrepancy suggests unresolved variability or a setup difference between the two experiments. With no seeds and no error bars in any of the training curves, the claimed quantum advantage may be within run-to-run noise. I ask for multiple independent runs per configuration (at least 3-5 seeds) with mean and standard deviation, and a reconciliation of the two classical baselines.","section":"Section IV.B, Figs. 7 and 8"},{"comment":"The channel-count arithmetic is not consistent. Section III.E.1 defines Cout = ncircuits x nqubits, and the circuit in Fig. 1 uses nqubits = 6, so the output channel count should be a multiple of 6. However, Fig. 3 lists quantum 1x1 layers with 64 and 128 output channels, neither of which is divisible by 6. The values of nqubits and ncircuits used in the experiments are never stated. This must be specified precisely, otherwise the architecture is under-specified and the reported parameter and accuracy comparisons cannot be reproduced.","section":"Section III.E and Fig. 3"},{"comment":"The conclusion claims 'significantly fewer parameters' for the quantum model, but no parameter count is reported for either model anywhere in the manuscript. The claim may be true (e.g., if the quantum circuits contain only a few dozen rotation angles), but as written it is unsupported. Please report the exact number of trainable parameters for the quantum and classical models, including the classical fully connected layers, and show the comparison explicitly.","section":"Section V, Conclusion"}],"minor_comments":[{"comment":"The text refers to 'multi-label classification' for FashionMNIST and CIFAR10, but these are multi-class classification tasks with mutually exclusive labels; the terminology should be corrected.","section":"Section IV.A"},{"comment":"The sentence 'Each qubit's state is measured using Pauli-Z operators transforming it into a quantum state' is garbled; measurement yields classical expectation values, not a quantum state.","section":"Fig. 1 caption"},{"comment":"The caption contains a typo: 'traning' should be 'training'.","section":"Fig. 4 caption"},{"comment":"Reference [20] is formatted incorrectly ('e, Kaiming'); this is the ResNet paper by He et al. and should be cited properly.","section":"Reference [20]"},{"comment":"The paper emphasizes NISQ deployability, but all experiments use the noiseless 'default.qubit' simulator; a brief discussion of noise sensitivity, or a small noise-model experiment, would clarify the practical claims.","section":"Section IV"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is not reproducible as submitted: no code, no seeds, and no hyperparameter details per model are given. Given that the main claim rests on a single uncontrolled experiment, I would ask the editor to treat the requested controlled comparison as a mandatory revision rather than a stylistic change."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper proposes quantum pointwise convolution—using a variational quantum circuit as a 1x1 kernel that maps each pixel's channel vector to a fixed number of output channels. That is a genuinely new variant: prior QCNNs operate over spatial patches, and this per-channel operation appears in no cited reference. The writing is clear, the PennyLane implementation is standard, and the discussion of the hybrid CPU-GPU bottleneck is honest. Those are real strengths.\n\nThe problem is that everything the paper claims about performance rests on one uncontrolled comparison. The classical baseline in Fig. 3b has BatchNorm and ReLU after every layer; the quantum model in Fig. 3a has no BatchNorm and a single ReLU. That architectural mismatch changes the optimization landscape and generalization behavior, so the reported accuracy gaps (quantum ~90% vs classical ~80% on CIFAR10) could come entirely from the missing normalization. There are no seeds, no error bars, no hyperparameter sweeps, and no code, so the comparison cannot be reproduced or evaluated. The classical model's accuracy is also internally inconsistent: it peaks near 80% in Fig 7 but near 70% in Fig 8 for the same architecture. The channel-count arithmetic does not reconcile either—the text says a circuit with n_qubits=6 produces six feature maps, so 64 and 128 output channels require non-integer numbers of circuits, and the paper never explains how that is handled. Finally, the conclusion's headline claim—\"significantly fewer parameters\"—is never backed by a single parameter count, which is doubly frustrating because that is the paper's main advertised advantage. The abstract says \"competitive performance\"; the conclusion says \"superior\"; neither is supported as written.\n\nI agree with the stress-test note. The central empirical claim is unsubstantiated. That said, the core idea is not bad, and a proper study with matched baselines, multiple seeds, and reported parameter counts could be worth publishing. As it stands, I would not cite this paper and would not send it to a top venue. If it crossed my desk, I'd reject with a clear path to revision rather than desk-reject out of hand, but only because the architecture is novel enough to merit one more chance with real experimental discipline.","headline":"Novel architecture, but the empirical comparison is too uncontrolled to support the claimed quantum advantage.","tokens_in":10167,"tokens_out":4007,"would_cite":false,"duration_ms":34716,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a parameterized quantum circuit can replace the pointwise (1x1) convolution layer in a CNN, matching or beating classical accuracy on FashionMNIST and CIFAR10 while using significantly fewer trainable parameters.","keywords":["quantum pointwise convolution","1x1 convolution","amplitude encoding","strongly entangling circuits","hybrid quantum-classical model","FashionMNIST","CIFAR10","parameter-shift rule"],"falsifier":"Train the classical network of Fig. 3(b) with BatchNorm removed to match the quantum network of Fig. 3(a), using the same random seeds, optimizer, and learning-rate schedule on CIFAR10; if this modified classical model reaches or exceeds roughly 90% accuracy, the claimed advantage of the quantum layer disappears.","tokens_in":9165,"feed_emoji":"⚛️","tokens_out":10562,"duration_ms":78594,"temperature":0.7,"pith_summary":"The paper proposes an architecture called Quantum Pointwise Convolution, in which the classical 1x1 pointwise convolution of a CNN is replaced by a parameterized quantum circuit. For each pixel, the vector of channel values is normalized and amplitude-encoded into a small register of qubits; the circuit then applies alternating rotations and entangling gates, and each qubit is measured to produce one output feature map. Because one circuit produces several feature maps and its parameters are shared across all pixels, the layer needs few trainable parameters. The authors report that in classification experiments on FashionMNIST and CIFAR10, the quantum version matches or outperforms an otherwise similar classical CNN while using significantly fewer parameters.","feed_headline":"Quantum 1x1 layer matches or beats classical net with fewer parameters","feed_subtitle":"A quantum pointwise convolution matches classical accuracy on FashionMNIST and CIFAR10 with far fewer parameters.","key_machinery":"The object that carries the argument is the quantum pointwise convolution kernel, a strongly entangling parameterized circuit. Its unitary is $U(\\theta) = \\prod_{\\ell=1}^L B_\\ell(\\theta_\\ell)$, where each block $B_\\ell$ alternates single-qubit rotations (Rz-Rx-Rz) with controlled phase/CNOT gates that entangle all qubits. Amplitude encoding compresses the channel vector into the state amplitudes, and measuring each qubit with a Pauli-Z operator turns one circuit into several output feature maps. Sharing the same rotation angles across every pixel gives the weight-sharing property of a classical convolution kernel, while the parameter-shift rule supplies the gradients that make the whole network trainable.","core_discovery":"The central claim is that pointwise convolution, normally a linear channel-mixing operation, can be implemented as a nonlinear quantum feature map without sacrificing accuracy. A single quantum circuit takes a pixel's channel vector, amplitude-encodes it into $n$ qubits, applies $L$ strongly entangling blocks of single-qubit rotations and CNOT gates, and reads out the Pauli-Z expectation of every qubit, yielding $n$ feature maps per circuit. Multiple circuits act as multiple convolution kernels, and circuit weights are shared over every spatial position. Trained end-to-end by the parameter-shift rule, this layer lifts the test accuracy of the demonstration network above 95% on FashionMNIST and to roughly 90% on CIFAR10, above the classical comparison, with fewer trainable parameters. In the paper's own words, substituting classical convolution with quantum circuits gave equivalent or better performance with significantly fewer parameters.","pith_inferences":["The comparison is not controlled: the classical model has BatchNorm and ReLU after every layer while the quantum model has neither, so the reported accuracy gap may reflect regularization differences rather than quantum feature maps.","The experiments run only on a noiseless simulator; on actual quantum hardware, amplitude encoding and entangling gates introduce noise that could erase the parameter-count advantage.","Since circuit parameters are shared across pixels and the circuit sees only the channel vector at one position, the layer is translation equivariant but spatially blind; its usefulness likely depends on neighboring spatial convolutions providing context.","A direct test suggested by the claim would be to insert the quantum layer into MobileNet's depthwise separable convolutions with matched BatchNorm on both sides, and compare accuracy and parameter counts on a larger dataset."],"forward_implications":["Wherever a CNN uses a 1x1 pointwise convolution, the quantum layer can be substituted directly, for example inside MobileNet-style depthwise separable blocks or ResNet bottleneck layers.","Because one circuit with $n$ qubits outputs $n$ feature maps, the layer expands channel counts while adding comparatively few trainable parameters.","Deepening the quantum circuit (up to four entangling layers in the experiments) monotonically improves accuracy, indicating that circuit depth is a useful resource for this layer.","Amplitude encoding keeps the qubit count logarithmic in the number of input channels, so the layer remains compact when channel counts grow."],"supporting_citations":[{"why":"Defines pointwise (1x1) convolution and its role in depthwise separable networks, the classical operation the quantum layer is designed to replace.","marker":"[11]"},{"why":"Supplies the amplitude encoding method that embeds the pixel's channel vector into the qubit register.","marker":"[16]"},{"why":"Provides the simulator and automatic differentiation tooling used to construct and train the quantum circuits.","marker":"[17]"},{"why":"Supplies the strongly entangling circuit-centric classifier design used as the quantum kernel.","marker":"[18]"},{"why":"Provides the parameter-shift rule used to compute gradients of circuit parameters during end-to-end training.","marker":"[19]"}],"fun_headline_variants":["Quantum pointwise convolution beats classical net on FashionMNIST and CIFAR10","Quantum 1x1 conv rivals classical with fewer parameters","Quantum pointwise conv: more accuracy, fewer parameters","Quantum circuits make 1x1 convolution accurate and compact"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the classical baseline, which has BatchNorm and ReLU after every layer, is a fair and adequately tuned comparison to the quantum model, which has no BatchNorm and only one ReLU.","fun_headline_variants_meta":{"raw":{"variants":["Quantum pointwise convolution beats classical net on FashionMNIST and CIFAR10","Quantum 1x1 conv rivals classical with fewer parameters","Quantum pointwise conv: more accuracy, fewer parameters","Quantum circuits make 1x1 convolution accurate and compact"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000272,"raw_usage":{"total_tokens":1624,"prompt_tokens":928,"completion_tokens":696,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":625}},"tokens_in":544,"tokens_out":696,"duration_ms":6437,"temperature":1.0,"reasoning_tokens":625,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:32:06.793115+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the classical network of Fig. 3(b) with BatchNorm removed to match the quantum network of Fig. 3(a), using the same random seeds, optimizer, and learning-rate schedule on CIFAR10; if this modified classical model reaches or exceeds roughly 90% accuracy, the claimed advantage of the quantum layer disappears.","supporting_citations":[{"cited_title":"Circuit-centric quantum classifiers,","cited_arxiv_id":null,"evidence_quote":"Supplies the strongly entangling circuit-centric classifier design used as the quantum kernel."},{"cited_title":"Wierichs, J","cited_arxiv_id":null,"evidence_quote":"Provides the parameter-shift rule used to compute gradients of circuit parameters during end-to-end training."}],"review_version":1}