{"id":"a4090df0-1a53-48f9-9d50-3a8e5c99b518","arxiv_id":"2504.18103","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Bayesian training of orthogonal quantum neural networks improves calibration for 3D anomaly detection, and an 8-qubit hardware test shows the pipeline tolerates device noise.","lead":"This paper combines Bayesian training with quantum-inspired orthogonal neural networks to detect defects in 3D-printed parts from CT scans. The approach yields better calibrated uncertainty estimates, and the authors test a small version on IBM quantum hardware.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline ECE advantage is not evaluable because the paper never defines how a continuous autoencoder reconstruction error becomes the predicted probabilities p_i required by Eq.","rationale":"The paper's strongest contribution is the empirical claim that Bayesian training improves ECE over point estimates in both feedforward and 3D-convolutional orthogonal networks, with a small IBM Brisbane hardware demonstration for the 3D-QCNN. The hardware fidelity experiments (Fig. 8) and the pipeline robustness experiment (Fig. 9) are concrete and useful, and the reported MSE values show a sensible error budget. The central calibration claim, however, cannot be assessed from the manuscript alone. Eq. (7) requires per-item predicted probabilities p_i, but the anomaly-detection pipeline in Section 4.2 outputs a reconstruction and a binary decision threshold; the mapping from reconstruction error to p_i is never given. This is not a stylistic omission: ECE is defined by those probabilities, so different mappings can change the ranking of models, and the paper's Table 1 caption's reference to a 'classification loss function' only adds ambiguity. The reader's weakest assumption identifies the same gap, and my independent reading confirms it. The concern is fixable by disclosure or a sensitivity analysis rather than by reworking the method, so it does not change the reader's CONDITIONAL verdict. I would not strengthen it to REJECT because the claimed effect is plausible and the missing pieces are specified enough to be supplied by the authors.","tokens_in":16817,"tokens_out":6480,"duration_ms":69070,"concrete_test":"Re-run the evaluation with an explicit, shared decision rule: (i) split at object level and train only on blocks whose masks are all zero; (ii) define the anomaly score as per-block MSE between input and reconstruction; (iii) set the binary threshold by a validation criterion such as maximum F1; (iv) map scores to probabilities with the same calibration function for every model, e.g., Platt scaling, and compute ECE with M=10 bins; (v) repeat with two alternative mappings (sigmoid((tau - MSE)/T) and rank-normalized MSE). If Bayesian models no longer beat point-estimate models on ECE under any stated mapping, the central calibration claim fails; if the advantage persists across mappings and train/test compositions, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.2 specifies only the reconstruction-error heuristic: D(E(X)) ≈ X means non-anomalous, otherwise anomalous. Section 2.7.1 defines ECE in Eq. (6) using bin confidences that average per-item predicted probabilities p_i (Eq. (7)). Nowhere does the paper state what p_i is for an autoencoder anomaly detector: reconstruction error is a continuous scalar, not a probability, and no sigmoid, temperature, Platt, or threshold-based calibration is described. Point-estimate models in particular have no natural probability over reconstruction error, so their ECE entries in Tables 1 and 2 are not well-defined; the Table 1 caption even mentions a 'classification loss function' although no classifier appears in the described pipeline. Independent of the probability mapping, the paper also does not state whether the autoencoder was trained only on non-anomalous 16×16×16 blocks or how train/test blocks are split relative to the ~5k scans. If anomalous blocks are present in training, the premise that anomalies are outliers with large reconstruction error is weakened; if blocks from the same scan appear in both train and test, metrics are optimistically correlated. These omissions are not cosmetic: they determine every number in the headline ECE comparison. A fully specified pipeline may well preserve the reported ranking, but the current text does not allow an independent check.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Bayesian learning of orthogonal (quantum) neural networks for anomaly detection in 3D CT scans of additively manufactured parts. The authors introduce OrthoConv3D, an orthogonal 3D convolution layer implemented with Hamming-weight-preserving circuits, embed it in an autoencoder, and compare Bayesian training against point-estimate training, Monte Carlo Dropout, and ensembling. They report metrics including precision, recall, F1, expected calibration error (ECE), and anomaly-size measures, and they present hardware experiments on IBM's 127-qubit Brisbane device showing that executing increasing fractions of the quantum circuits in the pipeline keeps the output MSE small.","tokens_in":17048,"tokens_out":5449,"duration_ms":51938,"significance":"If the evaluation pipeline were fully specified, the paper would make a useful contribution by combining Bayesian learning with orthogonal parameterizations for a practical anomaly detection task, and by demonstrating the architecture on real quantum hardware. The paper introduces a concrete architectural variant (OrthoConv3D) and uses a domain-relevant dataset with defect masks, which are strengths. However, the central comparison—the claimed ECE advantage of Bayesian methods—is not currently reproducible because the mapping from reconstruction error to the probabilities required by the ECE definition is never given, and the anomaly threshold and train/test split are unspecified. The hardware experiments are a useful feasibility data point but are limited to a single slice with no repeated runs or error bars. These issues block the central claims as written, though they appear fixable in a revision.","major_comments":[{"comment":"The ECE values in Tables 1 and 2 and Figures 3, 4, and 6 are not well-defined because the paper never specifies how the continuous autoencoder reconstruction error is converted to the per-item predicted probabilities p_i required by Eq. (7). Section 4.2 defines only the heuristic that a good reconstruction implies non-anomalous and a bad reconstruction implies anomalous, while Section 2.7.1 uses p_i without connecting it to the model output. For point-estimate models, which have no natural probability over reconstruction error, no sigmoid, temperature scaling, Platt scaling, or threshold-based probability map is described. The headline claim that \"Bayesian methods are clearly superior when ECE is the metric of interest\" (Table 1 caption) is therefore not independently checkable, and the caption's reference to a \"classification loss function\" has no counterpart in the described pipeline.","section":"2.7.1 and 4.2"},{"comment":"The anomaly detection metrics are underdetermined by two unspecified choices. First, reconstruction error is a continuous scalar, but no threshold is given for declaring a block anomalous; precision, recall, F1, SDA, and LuDA all depend on that threshold. Second, the paper does not state whether the autoencoder is trained only on non-anomalous 16×16×16 blocks, nor how the train/test split is made relative to the roughly 5k scans. If anomalous blocks enter training, the outlier premise is weakened; if blocks from the same scan appear in both train and test, the metrics are optimistically correlated. These choices determine every number in Tables 1 and 2, so the comparisons cannot be reproduced from the text.","section":"4.1-4.2"},{"comment":"All reported metrics are single point estimates without error bars, confidence intervals, or significance tests, although the differences are small (e.g., ECE 0.209 vs 0.224 for Bayesian 3D-CNN vs 3D-QCNN in Table 2). The claimed Bayesian advantage and any orthogonality advantage are therefore not statistically supported. In addition, Table 2 shows the 3D-QCNN is worse than the 3D-CNN on every listed metric, which contradicts any general claim that orthogonality improves anomaly detection; the discussion acknowledges this, but the abstract and the Table 1 caption should be qualified to the feedforward setting only.","section":"Tables 1-2 and Figures 3, 4, 6"},{"comment":"The hardware experiments support only a narrow conclusion. Figure 9 uses a single anomalous slice (slice 14 of one block), and the MSE values in the full-pipeline panel (c) come from one run, with no repeated measurements or variance estimate; the reported 1e-8 MSE may reflect the particular input. The fidelity estimate in Figure 8, eF = (ey·y)^2 with ey derived from estimated probabilities, is a heuristic rather than a standard fidelity estimator, and no error bars are shown for the device runs. The conclusion that quantum hardware can be incorporated into the pipeline would be more convincing with multiple inputs, repeated runs, and a description of how hardware error propagates to the final anomaly decision.","section":"5.1-5.2"}],"minor_comments":[{"comment":"The subplot letters in the captions do not match the labels in the figures: Figure 6 lists \"(c) Precision, (d) Recall, (e) F1-Score\" while the panels are labeled (d) through (f), and Figure 3 labels the SDA panel as the \"smallest undetected anomaly\" instead of the \"smallest detected anomaly.\"","section":"Figures 3 and 6 captions"},{"comment":"The sentence \"we take an anomalous voxel and downscale to 16×16×16\" should read \"an anomalous voxel block\" or \"a 16×16×16 volume,\" since a single voxel is a scalar.","section":"Section 5.2"},{"comment":"The construction of the rectangular filter matrix F_l from k filters of size d^3 is not fully specified; in particular, the paragraph should state how the case k ≠ d^3 is handled when using the square ortholinear circuit parameterization, beyond noting the max(k, d^3) qubit requirement.","section":"Section 3.1"},{"comment":"The reference list contains duplicate entries for the same paper (BCB+25a/25b and GPSM17a/17b); these should be consolidated.","section":"References"},{"comment":"The RBS gate matrix should be checked for sign conventions; as written, the (2,3) and (3,2) entries are consistent with the generator, but the ordering of basis states in the 4×4 matrix should be stated explicitly.","section":"Eq. (1)"}],"recommendation":"major_revision","confidential_remarks":"I recommend major revision rather than rejection because the missing pipeline details are fixable in a revision: defining the probability mapping for ECE, the anomaly threshold, and the data split would allow the central comparison to be evaluated. If those specifications reveal that the ECE ranking depends heavily on arbitrary threshold choices, the authors should downgrade the headline claim accordingly. The paper also needs error bars or significance tests to support its rankings."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this is a reasonable empirical study, not a breakthrough. The genuinely new pieces are modest: a 3D orthogonal convolution layer built from the known trick of expressing convolution as matrix multiplication, and mean-field Bayesian variational inference applied to the rotation angles of orthogonal quantum layers. That specific combination hasn't been published before, and it is a natural extension of the existing orthogonal QNN line. The authors also deserve credit for the hardware experiment: an 8-qubit fidelity test and a full-pipeline comparison on IBM Brisbane, including the honest observation that the full autoencoder output barely moves (MSE about 1e-8) even when all 256 circuits run on hardware. They also explicitly admit that orthogonality doesn't help in the 3D architecture, which matches Table 2.\n\nThe problem is the central calibration claim. ECE is defined in Eqs. (6)-(7) using per-item predicted probabilities p_i. The autoencoder anomaly detector produces a continuous reconstruction error, and nowhere in Section 4.2 or elsewhere is it explained how that scalar becomes p_i. No sigmoid, temperature, Platt scaling, or threshold-based calibration is described. So the ECE numbers in Tables 1 and 2 are not well-defined for any of the models, point-estimate or Bayesian. The Table 1 caption even mentions a 'classification loss function' even though no classifier appears in the described pipeline. This is not a cosmetic omission; it determines every number in the headline comparison. A fully specified mapping might preserve the ranking, but as written an independent check is impossible.\n\nTwo more soft spots, in decreasing seriousness. First, there are no error bars, confidence intervals, or significance tests anywhere. The ECE differences are small (e.g., 0.257 vs 0.221) and could easily be noise. Second, we never learn whether the autoencoder was trained only on normal blocks, or whether blocks from the same CT scan appear in both train and test. Both choices affect the anomaly detection premise and every metric. The dataset is proprietary and no code is released, so reproducibility is limited.\n\nWhat holds up: the circuit construction, the 3D convolution reformulation, and the Bayesian training objective are standard and correctly described. The paper doesn't overclaim scalability; it calls the hardware part a feasibility test. The citation pattern is fine. The main issue is missing specification, not circular reasoning or bad faith.\n\nWho this is for: researchers working on quantum or quantum-inspired orthogonal networks, or on uncertainty quantification in QML. It deserves peer review, but it needs a substantial revision to specify the full anomaly-detection pipeline, add error bars, and soften the 'clearly superior' language. I'd send it to a serious venue but expect heavy revision.","headline":"Solid but incremental engineering story; the headline ECE advantage is not independently checkable because the paper never defines how reconstruction error becomes the predicted probabilities that ECE requires.","tokens_in":17556,"tokens_out":3334,"would_cite":false,"duration_ms":33635,"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":"The paper claims that Bayesian learning of rotation angles in orthogonal quantum neural networks gives anomaly-detection models whose confidence is better calibrated than point-estimate training, and that this holds in an autoencoder…","keywords":["Bayesian learning","orthogonal neural networks","quantum machine learning","anomaly detection","3D convolutional networks","uncertainty calibration","variational inference","additive manufacturing"],"falsifier":"Re-run the four model families with a publicly specified reconstruction-error threshold and a fixed mapping from reconstruction error to confidence, training only on voxel blocks verified to be anomaly-free; if Bayesian ECE no longer beats point-estimate ECE under that exact protocol, the paper's central calibration claim is refuted.","tokens_in":16628,"feed_emoji":"⚛️","tokens_out":7171,"duration_ms":63560,"temperature":0.7,"pith_summary":"The paper tries to establish that training the rotation angles of orthogonal quantum neural networks as probability distributions, instead of as point estimates, yields anomaly-detection models whose confidence is better calibrated. In the autoencoder pipeline applied to 3D voxel blocks from CT scans, Bayesian training lowers the expected calibration error for both fully connected and the paper's new 3D-convolutional architectures, though at a small cost in precision, recall, and F1 score. A secondary claim is that the new OrthoConv3D layer, built from Hamming-weight-preserving quantum circuits, can be embedded in the pipeline, and that executing its circuits on a current 127-qubit processor keeps the full reconstructed object within about $10^{-8}$ mean squared error of the noiseless result. If right, this would make Bayesian quantum-inspired networks a practical route to uncertainty-aware defect detection in industrial 3D-printing quality control.","feed_headline":"Bayesian training cuts calibration error in quantum anomaly detection","feed_subtitle":"Rotation angles are learned as distributions, lowering expected calibration error in 3D-printed part inspection.","key_machinery":"The load-bearing object is the orthogonal layer built from reconfigurable beam splitter (RBS) gates, two-qubit rotations that preserve Hamming weight; input vectors are loaded into the unary subspace and the circuit implements an orthogonal matrix whose entries are trigonometric functions of the gate angles. Bayesian learning is applied directly to these angles: each angle is a Gaussian random variable with trainable mean and variance, optimised through the ELBO, so sampling the angles gives a distribution over orthogonal matrices. OrthoConv3D flattens each $d\\times d\\times d$ patch and each $k\\times d^3$ kernel matrix, implements the kernel matrix as such a circuit, and multiplies it by the patch vectors. The expected calibration error (ECE) is the metric that carries the comparison, since it measures whether the model's stated confidence matches its accuracy across probability bins.","core_discovery":"The paper's central claim is that treating the rotation angles of orthogonal quantum neural-network layers as random variables, with means and variances learned by variational inference, produces models whose confidence is better calibrated than identical architectures trained by point-estimate gradient descent. In the autoencoder anomaly-detection pipeline, Bayesian training lowers the expected calibration error (ECE) for both fully connected networks and the new 3D orthogonal convolutional architecture, at the price of slightly lower precision, recall, and F1 score. The paper also introduces OrthoConv3D, a 3D convolutional layer whose flattened kernel matrix is enforced to be orthogonal via Hamming-weight-preserving quantum circuits, and reports that executing all 256 component circuits on quantum hardware leaves the full reconstructed object within roughly $10^{-8}$ mean squared error of the noiseless simulation. Orthogonality helps the feedforward models, giving competitive metrics with fewer parameters, but it does not beat classical 3D convolutions on the standard detection metrics.","pith_inferences":["The reported ECE advantage is only as meaningful as the unstated mapping from autoencoder reconstruction error to anomaly labels and confidence probabilities; a reader cannot currently reproduce the calibration numbers from the text alone.","If the training set contains any anomalous voxel blocks, the autoencoder's premise that anomalies reconstruct poorly is weakened, so the SDA and LuDA comparisons may partly reflect memorisation rather than outlier detection.","The 8-qubit hardware demonstrations leave open whether larger orthogonal kernels retain the same noise tolerance; the linear-depth circuit construction suggests it could, but the paper does not test this."],"forward_implications":["Bayesian training lowers ECE relative to point-estimate training across all four model families considered, so a deployment that prioritises calibrated confidence should prefer the Bayesian variant.","The orthogonal feedforward models achieve competitive detection metrics with fewer parameters than vanilla feedforward models, while the orthogonal 3D convolutions do not beat their classical counterparts on precision, recall, or F1.","The full autoencoder pipeline keeps reconstruction error near the noiseless value even when all 256 orthogonal circuits run on current quantum hardware, suggesting that device noise and limited shots are not blockers for this hybrid pipeline.","Because Bayesian training here trades away precision, recall, and F1 score, the choice between Bayesian and point-estimate training is a judgement about whether calibration or raw detection rate matters more for the application."],"supporting_citations":[{"why":"Supplies the quantum orthogonal layer construction via Hamming-weight-preserving RBS gates and the pyramid or butterfly parameterisations used throughout.","marker":"[LML+22]"},{"why":"Provides the butterfly layout and the quantum vision transformer framework from which the orthogonal parametrisation is drawn.","marker":"[CKM+22]"},{"why":"Reformulates convolution as matrix multiplication, the basis for the OrthoConv3D construction.","marker":"[KLP19]"},{"why":"Supplies the loader circuit that encodes flattened patches into unary states.","marker":"[JDM+21]"},{"why":"Grounds Monte Carlo Dropout as an approximate Bayesian method used as a comparison baseline.","marker":"[GG16]"},{"why":"Grounds the ensembling baseline for uncertainty estimation.","marker":"[LPB17]"},{"why":"Motivates Bayesian learning of parameterised quantum circuits and the variational posterior framework used here.","marker":"[DBR23]"},{"why":"Provides the classical orthogonal convolution idea that OrthoConv3D extends to three dimensions.","marker":"[WCCY20]"}],"fun_headline_variants":["Bayesian quantum nets reduce calibration error in 3D anomaly checks","Quantum orthogonal nets with Bayes: better confidence","Calibration boost from Bayesian quantum 3D anomaly nets","Bayes and orthogonality improve 3D defect detection confidence"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported calibration advantage rests on an unstated mapping from autoencoder reconstruction error to binary anomaly labels and to confidence probabilities for ECE, and on the unstated assumption that training blocks contain only normal voxels; the numbers in the tables stand or fall with those choices.","fun_headline_variants_meta":{"raw":{"variants":["Bayesian quantum nets reduce calibration error in 3D anomaly checks","Quantum orthogonal nets with Bayes: better confidence","Calibration boost from Bayesian quantum 3D anomaly nets","Bayes and orthogonality improve 3D defect detection confidence"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000581,"raw_usage":{"total_tokens":2688,"prompt_tokens":853,"completion_tokens":1835,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":469,"completion_tokens_details":{"reasoning_tokens":1767}},"tokens_in":469,"tokens_out":1835,"duration_ms":14119,"temperature":1.0,"reasoning_tokens":1767,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:23:37.187164+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the four model families with a publicly specified reconstruction-error threshold and a fixed mapping from reconstruction error to confidence, training only on voxel blocks verified to be anomaly-free; if Bayesian ECE no longer beats point-estimate ECE under that exact protocol, the paper's central calibration claim is refuted.","supporting_citations":[],"review_version":1}