{"id":"87a3dbd0-a845-4e0e-ad22-a855159919ca","arxiv_id":"2411.13468","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A simplified real-valued quantum convolutional neural network matches the classification accuracy of a hardware-efficient ansatz while training roughly ten times faster.","lead":"This paper benchmarks quantum convolutional neural networks against hardware-efficient ansatze for classifying quantum ground states and for compressing quantum states. It reports that a simplified real-valued QCNN variant trains much faster while matching the accuracy of a much larger ansatz on the tested tasks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline speed/accuracy comparison is not statistically grounded: a single unseeded run with no error bars and unmatched training budgets leaves 'matching performance' and the 10x speedup unestablished.","rationale":"I read the paper as a benchmark study whose central claim is that a 17-parameter RY-gate QCNN reaches HEA-level accuracy approximately 10x faster. For that claim to hold, the accuracy comparison must be statistically meaningful and the timing comparison must be apples-to-apples. The manuscript provides neither repeated trials with error bars nor a description of optimization budgets or stopping criteria; the Table I gap of 0.931 vs 0.938 is plausibly noise. The reader's identified concern about the 16-qubit variational-circuits dataset is real and worth addressing, but it would mainly affect whether the benchmark is about true ground states, not whether the model comparison is internally reliable. I therefore partially agree with the reader; the same conditional verdict stands, with the added requirement for statistical and budget-matched evidence.","tokens_in":3912,"tokens_out":7297,"duration_ms":80843,"concrete_test":"Repeat the 16-qubit classification experiment for QCNN(RY) and HEA(3L) with at least 10 independent random initializations, identical optimizer settings, and a fixed convergence rule or fixed epoch budget. Report mean ± std test accuracy and total wall-clock time, not only per-sample time. If the mean accuracies are within one standard deviation and the speedup persists under matched budgets, the central claim is supported; if variance is large or the time gap shrinks, the claim needs qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table I and Fig. 2 support the central claim with a single unseeded comparison: QCNN(RY) test accuracy 0.931 versus HEA(3L) 0.938 at 60 training samples. No standard deviation, number of random initializations, or significance test is reported. QCNN and HEA training is stochastic; a 0.007 accuracy gap is likely within run-to-run variance for 16-qubit circuits, so 'matching performance' is not established. The 10x training-time advantage is also attributed to fewer parameters, but Fig. 2b reports 'training time per training sample' without specifying optimizer settings, convergence criteria, epoch counts, or whether QCNN and HEA were given equal optimization budgets. If the QCNN was trained for fewer epochs or stopped by a different rule, the time comparison is confounded with training budget rather than parameter count. The reader's concern about the 16-qubit dataset (optimized variational circuits rather than exact ground states) is valid but secondary: it affects external validity of the benchmark, not the internal reliability of the QCNN-versus-HEA comparison.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper benchmarks quantum convolutional neural networks (QCNNs) against hardware-efficient ansätze (HEAs) on two tasks: phase classification of transverse-field Ising (TFI) and XXZ ground states at 4, 8, and 16 qubits, and quantum autoencoder compression of TFI ground states. The authors propose real-valued QCNN variants, including an RY-gate version, and report that the 17-parameter QCNN(RY) reaches test accuracy 0.931 at 16 qubits for 60 training samples, close to the 113-parameter HEA(3L) accuracy of 0.938, while taking 18.7 s per training sample versus 188.7 s. They conclude that QCNNs with RY gates are the most parameter-efficient architecture and also converge faster in the autoencoder task. The manuscript reports no error bars or repeated-seed statistics, gives limited experimental detail for the training-time comparison, and includes only one quantitative figure/table in the visible text.","tokens_in":4096,"tokens_out":4706,"duration_ms":46282,"significance":"If the performance comparison were statistically and experimentally validated, the result would be practically useful: it would identify a low-parameter QCNN architecture that matches a higher-parameter HEA on phase classification while reducing training time, and it would provide evidence for QCNN advantages in quantum data compression. Strengths of the paper include the use of established simulation tools (Qibo), explicit parameter counts for all models, comparison against a standard HEA baseline, and clear statements of the cost functions. The paper does not claim a theoretical advance; its contribution is empirical. The main barrier to significance is the lack of statistical grounding and the unverified 16-qubit input states.","major_comments":[{"comment":"The claim that QCNN(RY) matches HEA(3L) performance rests on a single unseeded comparison (0.931 vs 0.938 at 60 training samples). Because quantum circuit training is stochastic, this 0.7 percentage point gap is plausibly within run-to-run variance; without error bars, multiple random initializations, or a significance test, 'matching performance' is not established. This is load-bearing for the paper's central conclusion.","section":"Sec. III, Fig. 2 and Table I"},{"comment":"The reported 10x training-time advantage is not grounded because the training time per sample is presented without specifying the optimizer settings, convergence criteria, or number of epochs, and without evidence that the QCNN and HEA runs were given equal optimization budgets. If the models were trained for different durations or stopped by different rules, the time comparison conflates parameter count with training budget. The authors should report wall-clock time to reach a fixed target accuracy, or per-epoch time under matched budgets, across multiple seeds.","section":"Sec. III, Fig. 2b and Table I"},{"comment":"The 16-qubit datasets are described as 'optimized variational circuits provided in the tensorflow-quantum repository' rather than exact ground states of the TFI and XXZ models. The paper does not verify how closely these circuits approximate the true ground states, so the 16-qubit accuracies in Fig. 2 and Table I may not correspond to the claimed phase-classification benchmark. This is a concern for external validity even if the internal QCNN-versus-HEA comparison remains meaningful.","section":"Sec. II.C"},{"comment":"The data-compression section contains no quantitative results: no reconstruction fidelities, learning curves, parameter counts, or training times are reported for the autoencoder experiments. The claims that QCNNs 'demonstrated faster training convergence' and exhibit 'minimal tradeoff in compression capability' are therefore unsupported in the manuscript as written.","section":"Sec. II.B and Sec. III"},{"comment":"The paper states that system sizes of 4, 8, and 16 qubits were examined, but Fig. 2 reports results only for 16 qubits; the 4- and 8-qubit classification results and the autoencoder results are not shown. To support the claimed scaling behavior and the comparison across system sizes, these results should be presented or the scope statement should be revised.","section":"Sec. I and Sec. III"}],"minor_comments":[{"comment":"There are formatting issues: Eqs. (1) and (2) render with broken spacing, the inner product in the label prediction is missing angle brackets, and 'nd = N (1 − 1/2l)' should read 'nd = N(1 − 1/2^l)'.","section":"Throughout"},{"comment":"The header says 'best QCNN and VQC models'; 'VQC' should likely be 'HEA' to match the text.","section":"Table I"},{"comment":"The phrase 'through simulation were examined' is grammatically incomplete and should be revised.","section":"Abstract"},{"comment":"The paper reports 'training time per training sample (s)' but does not define how this is measured; please clarify the measurement procedure, hardware, and software versions.","section":"Sec. II.C and Fig. 2b"},{"comment":"The statement that 'Powell is one of the most reliable optimizers' is not supported by any reported optimizer comparison; either add the comparison data or soften the claim.","section":"Sec. IV"},{"comment":"The exact architecture details of QCNN, QCNN(real), QCNN(RY), and HEA are not fully specified; a circuit diagram or explicit gate-layer list for each variant would improve reproducibility.","section":"Sec. II.A"}],"recommendation":"major_revision","confidential_remarks":"This is a straightforward empirical benchmark without a theoretical advance; it is likely best suited to an applied quantum-machine-learning venue. The main fixable weaknesses are statistical rigor and missing quantitative details for the autoencoder and small-system results. I support a major revision rather than rejection, because the reported parameter-count/training-time tradeoff is a useful observation if confirmed with proper statistical controls."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague—\n\nQuick take: this is a modest but honest benchmark paper. The useful bit is the real-valued QCNN variant: using RY gates only, 17 parameters, it gets test accuracy 0.931 vs 0.938 for a 113-parameter HEA(3L) at 16 qubits, and trains ~10x faster per sample. If that holds, it's a nice practical data point for QML. The paper also runs a QCNN/HEA autoencoder comparison and gets high reconstruction fidelities.\n\nThe real-valued QCNN is a small architectural tweak to Cong et al.'s QCNN, not a new framework, and the head-to-head benchmark against HEA is genuinely new. The work is careful about distinguishing expressibility and trainability, and the observation that real-valued ground states let you restrict to RY gates is reasonable. For 4 and 8 qubits the training data come from exact diagonalization, which is the right way to do it.\n\nNow the soft spots. The headline comparison is not statistically grounded. Table I and Fig. 2 rest on a single unseeded run; no standard deviations, no repeated initializations, no significance test. A 0.007 accuracy gap on 16-qubit circuits is easily within run-to-run variance, so 'matching performance' is not actually demonstrated. The 10x training-time claim is also confounded: without specifying optimizer settings, epoch counts, convergence criteria, or whether both models got the same optimization budget, part of the speedup could be a shorter budget rather than fewer parameters. That is the stress-test's concern and it holds up. The 16-qubit dataset from TensorFlow Quantum variational circuits is a second, real issue: those are not exact ground states, so the external validity of the 16-qubit comparison is weaker than the 4/8-qubit part. It is secondary to the internal QCNN-vs-HEA comparison, but it should be disclosed and ideally replaced or verified.\n\nThe missing architecture details and code are minor-to-moderate; enough is described to get the gist but not to reproduce without guessing. These are all fixable.\n\nOverall: the central claim is plausible but under-supported as published. The paper deserves a serious referee; I would send it to peer review with a demand for repeated seeds, error bars, matched training budgets, and clarification of the 16-qubit data. Useful for QML practitioners choosing ansätze.","headline":"Plausible benchmark showing a parameter-light RY-QCNN matches larger HEAs, but the headline speed/accuracy comparison lacks error bars and matched budgets.","tokens_in":4585,"tokens_out":2321,"would_cite":false,"duration_ms":25407,"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 17-parameter QCNN matches a 113-parameter ansatz at about one-tenth the training time per sample.","keywords":["quantum convolutional neural networks","hardware-efficient ansatz","phase classification","quantum autoencoder","transverse field Ising model","XXZ model","quantum machine learning","trainability"],"falsifier":"Recompute the 16-qubit classification using exact ground states (or high-fidelity tensor-network states) of the TFI and XXZ models at the same field values; if the RY-QCNN's accuracy falls below the HEA's once the labels are exact, the paper's flagship comparison fails.","tokens_in":3737,"feed_emoji":"⚛️","tokens_out":6381,"duration_ms":61449,"temperature":0.7,"pith_summary":"This paper benchmarks quantum convolutional neural networks (QCNNs) against hardware-efficient ansatze (HEAs) on two tasks: classifying ground-state phases of the transverse-field Ising and XXZ models, and compressing Ising ground states with quantum autoencoders. The central result is that a QCNN restricted to RY rotation gates, with only 17 trainable parameters, reaches essentially the same test accuracy as the best HEA (0.931 versus 0.938 at 16 qubits) while costing roughly one tenth of the training time per sample. More expressive QCNN variants perform worse, which the authors attribute to trainability problems rather than expressive power. For equal parameter counts, the RY-QCNN is the most efficient architecture on these benchmarks, and the same parameter-count advantage explains faster convergence in the compression task.","feed_headline":"QCNN with 17 parameters matches 113-parameter ansatz in 1/10 time","feed_subtitle":"Quantum benchmark on ground-state classification and compression shows fewer parameters train faster without losing accuracy.","key_machinery":"The central object is the QCNN circuit, whose convolutional and pooling layers halve the qubit count at each step, giving logarithmic depth and a structured reduction in trainable parameters; the RY-restricted variant keeps only real-valued rotation gates, cutting parameters from 85 (complex QCNN) to 17. The HEA comparator is a layered ansatz of single-qubit RY rotations and entangling gates whose parameter count grows with layers. The argument runs through parameter count: fewer parameters shrink the optimization landscape and per-sample cost, and the logarithmic-depth QCNN structure is the reason trainability is preserved. For compression, the autoencoder uses the encoder circuit's inverse as the decoder, with a cost function that drives the discarded qubits to $|0\\rangle^{\\otimes n_d}$.","core_discovery":"The paper claims that for phase classification of TFI and XXZ ground states on 4, 8, and 16 qubits, a QCNN built from RY gates matches the best-performing HEA in accuracy while needing far fewer parameters and much less training time. Its flagship comparison at 16 qubits is a 17-parameter RY-QCNN at 0.931 test accuracy and 18.7 seconds per training sample versus a 113-parameter three-layer HEA at 0.938 accuracy and 188.7 seconds per sample. The authors also claim that increasing QCNN expressibility beyond this real-valued form degrades performance, while HEAs improve with more layers at linearly increasing cost. In the autoencoder task, both architectures reconstruct TFI ground states with high fidelity, but QCNNs converge faster because of their smaller parameter counts.","pith_inferences":["If the training-time advantage survives on real hardware, the RY-QCNN's low parameter count makes it a practical default for near-term devices; the paper only reports simulation and lists hardware comparison as future work.","The 16-qubit result rests on variational proxy ground states rather than exact diagonalization, so the 16-qubit ranking should be read as provisional until exact-state labels confirm it.","The real-valued RY restriction is motivated by the fact that these ground states are real; on Hamiltonians whose ground states require complex phases, the RY-QCNN advantage may shrink or disappear.","A natural testable extension is to run the same benchmark on disordered or long-range models, where the QCNN's locality assumption may fail."],"forward_implications":["If the benchmark is representative, RY-QCNN achieves near-identical accuracy to the best HEA with a roughly tenfold reduction in per-sample training time at 16 qubits.","For equal trainable-parameter budgets, the RY-QCNN is the most efficient architecture for classifying these ground states.","The more expressive QCNN variants' poorer accuracy is a trainability effect, not a capacity limit, so architecture choice should weigh optimizability, not just expressibility.","On the compression task, QCNN-based autoencoders reconstruct TFI ground states with high fidelity while converging faster than HEA-based autoencoders."],"supporting_citations":[{"why":"Defines the QCNN architecture whose RY-restricted variant is the paper's best-performing model.","marker":"[1]"},{"why":"Establishes that QCNNs avoid barren plateaus, supporting the trainability explanation for the results.","marker":"[3]"},{"why":"Introduces the quantum autoencoder scheme with the inverse encoder as decoder and the discarded-qubit cost function.","marker":"[4]"},{"why":"Provides the GPU-accelerated quantum circuit simulator used for the benchmark simulations.","marker":"[6]"},{"why":"Supplies the 16-qubit labelled dataset as optimized variational states rather than exact ground states.","marker":"[7]"}],"fun_headline_variants":["Quantum CNN trains 10x faster with 1/6 the parameters","17-parameter QCNN matches 113-parameter ansatz in benchmarking","RY-QCNN: 10x faster training, HEA-level accuracy","QCNN benchmarking: fewer parameters, same accuracy, faster","Quantum CNN slashes parameters and training time without accuracy loss"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The 16-qubit dataset labels come from optimized variational circuits bundled with a quantum machine-learning software repository, not from exact ground states; if those circuits deviate from the true TFI and XXZ ground states, the 16-qubit accuracies compare classifiers on the wrong labels.","fun_headline_variants_meta":{"raw":{"variants":["Quantum CNN trains 10x faster with 1/6 the parameters","17-parameter QCNN matches 113-parameter ansatz in benchmarking","RY-QCNN: 10x faster training, HEA-level accuracy","QCNN benchmarking: fewer parameters, same accuracy, faster","Quantum CNN slashes parameters and training time without accuracy loss"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000944,"raw_usage":{"total_tokens":3976,"prompt_tokens":833,"completion_tokens":3143,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":449,"completion_tokens_details":{"reasoning_tokens":3053}},"tokens_in":449,"tokens_out":3143,"duration_ms":24879,"temperature":1.0,"reasoning_tokens":3053,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:22:02.422935+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the 16-qubit classification using exact ground states (or high-fidelity tensor-network states) of the TFI and XXZ models at the same field values; if the RY-QCNN's accuracy falls below the HEA's once the labels are exact, the paper's flagship comparison fails.","supporting_citations":[{"cited_title":"Quan- tum convolutional neural networks","cited_arxiv_id":null,"evidence_quote":"Defines the QCNN architecture whose RY-restricted variant is the paper's best-performing model."},{"cited_title":"Cerezo, Samson Wang, Tyler Volkoff, Andrew T","cited_arxiv_id":null,"evidence_quote":"Establishes that QCNNs avoid barren plateaus, supporting the trainability explanation for the results."},{"cited_title":"Quantum autoencoders for efficient compression of quantum data","cited_arxiv_id":null,"evidence_quote":"Introduces the quantum autoencoder scheme with the inverse encoder as decoder and the discarded-qubit cost function."},{"cited_title":"Qibo: a framework for quantum simulation with hardware acceleration","cited_arxiv_id":null,"evidence_quote":"Provides the GPU-accelerated quantum circuit simulator used for the benchmark simulations."},{"cited_title":"Martinez, Jae Hyeon Yoo, Sergei V","cited_arxiv_id":null,"evidence_quote":"Supplies the 16-qubit labelled dataset as optimized variational states rather than exact ground states."}],"review_version":1}