acceptodds
Desk rejected at ICLR 2027

Quantum Sharpness-Aware Minimization

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Quantum neural networks (QNNs) deployed on noisy intermediate-scale quantum (NISQ) devices often suffer from severe performance degradation due to intrinsic hardware noise. By spotlighting this challenge from a deep neural network (DNN) perspective, we raise an intriguing question: *Can we incorporate a concept of flat minima in DNNs into QNNs for quantum noise robustness?* As our answer, we propose quantum sharpness-aware minimization (Q-SAM), which integrates the flat minima concept into the Riemannian quantum parameter space to seek minima that remain *flat* or invariant under quantum state perturbations induced by hardware noise. We formally connect a bounded perturbation in the quantum state manifold to a bound in the Riemannian quantum parameter space, providing theoretical ground for Q-SAM's state noise robustness. Across image classification tasks, including MNIST-10, Fashion-MNIST, and CIFAR-10, Q-SAM with 4-, 6-, and 8-qubit QNNs consistently exhibit superior robustness against diverse noise channels compared to baselines, also validated on real IBM quantum hardware.

Primary area:
optimization
Keywords:
Quantum Machine Learning, Robustness

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_3eZhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper is trying the address the problem that when QNNs deploy on quantum devices, due to intrinsic hardware noise, it suffer from severe performance degradation. The paper incorporates a flat minimum in DNNS into QNNs for quantum noise robustness. It proposes quantum sharpness-aware minimization (Q-SAM), which integrates the flat minima concept into the Riemannian quantum parameter space to seek minima that remain flat or invariant under quantum state perturbations induced by hardware noise. The proposed method was tested across image classification tasks with 4, 6, 8-quibit QNNs and validated on IBM quantum hardware.

Critical strengths

The paper is addressing an important practical problem, the geometry-aware optimization approach with theoretical motivation and evaluation across multiple noise models and different noise combinations evaluations on real quantum hardware.

Critical weaknesses

The experimental evaluation primarily limited to image classification benchmarks. However, the proposed Q-SAM aims to improve the robustness of quantum neural networks against hardware noise. It would be valuable to evaluate it’s generalizability beyond image classifications. Table 1 shows different types of noise which can be applied on single qubit, I wonder if these 5 types of noise and noise combinations cover all the noises. Could the author further clarify whether the simulations incorporate hardware-calibrated noise models?

Questions for authors
  1. Equation (4) Defined the optimization objective, which is QNN-loss deviation |L(˜ρ) − L(ρ)|, but I cannot find the definition of L(ρ). Please explain how to define this function and why it is defined in this way.
  2. The paper uses IBM Heron(133-qubit) hardware, but only evaluates 4, 6, 8 qubit QNNs, I think larger qubit experiments might be useful. As the nuber of qubits increase, there may be some noise accumulation.
  3. From line 359-365, the paper tries different number of qubit on different image classification datasets, how about the computational cost when increasing the number of qubit?
  4. Section 5.2 shows evaluation on both single-source noise and on composite noise, given that the paper aims to improve the robustness against quantum hardware noise, I wonder if the authors have done some testing on cross noise generalization, which means training on certain types of noises and testing on noises that not injected in training data.
3 comments

We thank the reviewer for their feedback. Below, we address robustness beyond image classification, hardware noise, and circuit scaling.

W1. Beyond image classification and hardware noise


Beyond image classification. We added two 10-class tasks, spoken command recognition on Speech Commands [1] and text emotion classification on GoEmotions [2]. Prior QNN studies address spoken commands [3] and text sentiment [4].

As in Table 2, we test each noise channel separately. Section 5.2 (Table 4) reports all four optimizers, with setup details in Appendix D.4 (Table 15). Perturbed test accuracies are averaged over five seeds. Parentheses show changes from nominal accuracy in percentage points.

Speech Commands. We used a 4-qubit data-reuploading circuit based on our MNIST architecture. Q-SAM improves mean noisy accuracy by 7.13 percentage points across the five channels while preserving nominal accuracy (78.15% vs. 77.95%).

OptimizerNominalDepolarizing (0.0668)Amplitude damping (0.222)Phase damping (0.5832)Readout bit flip (0.31)Phase flip (0.1753)
Adam77.9535.70 (−42.24)21.68 (−56.27)40.18 (−37.77)39.29 (−38.66)40.47 (−37.48)
Q-SAM78.1541.58 (−36.58)29.23 (−48.92)49.11 (−29.05)43.55 (−34.60)49.49 (−28.66)

GoEmotions. We used the 8-qubit circuit from our CIFAR-10 experiment. Q-SAM improves mean noisy accuracy by 14.65 percentage points, with a 3.14-point decrease in nominal accuracy.

OptimizerNominalDepolarizing (0.025)Amplitude damping (0.12)Phase damping (0.36)Readout bit flip (0.27)Phase flip (0.015)
Adam53.6912.51 (−41.17)29.00 (−24.69)16.63 (−37.06)18.31 (−35.37)16.72 (−36.97)
Q-SAM50.5523.45 (−27.09)37.83 (−12.71)32.78 (−17.76)39.91 (−10.64)32.44 (−18.10)

Hardware noise. The five single-qubit channels in Table 1 are widely accepted as common hardware errors. The table below shows how these errors cover common erroneous physical cases:

Noise channelHardware errorChannel parameter
DepolarizingStochastic gate errorp=1−1−5rCZ/4p=1-\sqrt{1-5r_{\mathrm{CZ}}/4}
Amplitude dampingEnergy relaxationγAD=1−e−t/T1\gamma_{\mathrm{AD}}=1-e^{-t/T_1}
Phase dampingPure dephasingγPD=1−e−2t/Tϕ\gamma_{\mathrm{PD}}=1-e^{-2t/T_\phi}, Tϕ−1=T2−1−(2T1)−1T_\phi^{-1}=T_2^{-1}-(2T_1)^{-1}
Phase flipStochastic phase errorspZ=1−e−t/Tϕ2p_Z=\frac{1-e^{-t/T_\phi}}{2}
Readout bit flipSymmetric assignment errorpX=P(1∣0)+P(0∣1)2p_X=\frac{P(1\mid0)+P(0\mid1)}{2}

Applying these local channels across qubits and layers allows multiple errors to occur in the same circuit and propagate through entangling gates. Qiskit Aer also combines gate, relaxation, and readout errors in its device noise models [5].

Composite noise and Heron calibration. Section 5.2 and Table 3 already include a composite-noise simulation to approximate IBM Heron errors. We apply gate depolarization, thermal relaxation, and readout error together during inference. McKay et al. [6, Appendix B] describe combining gate errors with relaxation over circuit time steps. The table below compares our settings with the corresponding ranges from IBM’s archived ibm_torino calibration data [7].

Error mechanismTable 3 simulationIBM Heron range [7]
Gate error0.01 total Pauli error probability0.00083–0.02043 (equivalent Pauli error probability)
Relaxation and dephasing0.005 layer-level relaxation strength0.00027–0.00992 (amplitude damping), 0–0.02460 (phase damping)
Readout error2.0% symmetric bit flip0.51–18.96% (median 1.96%)

Table 3 uses representative noise strengths within these Heron calibration ranges.

In Table 3, Q-SAM achieved 44.16% vs. 29.66% for Adam under composite noise and 49.33% vs. 27.33% on the physical Heron processor.

References

[1] Warden. Speech Commands. 2018.

[2] Demszky et al. GoEmotions. ACL 2020.

[3] Qi and Tejedor. Classical-to-Quantum Transfer Learning for Spoken Command Recognition Based on Quantum Neural Networks. 2021.

[4] Liu et al. HCQNSC. EPJ Quantum Technology, 2026.

[5] Qiskit Aer. Device backend noise model simulations.

[6] McKay et al. Benchmarking Quantum Processor Performance at Scale.

[7] ibm_torino calibration snapshot (28 February 2024)

Q1. Definition of the QNN loss


Appendix A.2 already defines L(ρ)L(\rho) using the measurement in Eq. (31). For a circuit output state ρ\rho,

Lmeas(ρ)=Tr⁡(Oρ),L(ρ)=ℓ(Lmeas(ρ)).L_{\mathrm{meas}}(\rho)=\operatorname{Tr}(O\rho),\qquad L(\rho)=\ell\left(L_{\mathrm{meas}}(\rho)\right).

Here, OO denotes the measured observables and ℓ\ell applies the classical post-processing and cross-entropy loss for a fixed target label. Noise can change the measured values and, in turn, the classification loss. The quantity ∣L(ρ~)−L(ρ)∣\lvert L(\tilde{\rho})-L(\rho)\rvert measures this change for the same input and label.

We added this definition immediately before Eq. (4).

Q2. Larger circuits and noise accumulation


Appendix D.2 already includes a 10-qubit experiment using amplitude encoding of 28×2828\times28 inputs. Q-SAM achieves the highest noisy accuracy across all five channels, with gains of up to 11.40 percentage points over the strongest baseline.

OptimizerNominalDepolarizingAmplitude dampingPhase dampingBit flipPhase flip
Adam90.9247.8849.5688.8341.3888.34
SAM91.1252.3452.8288.3146.5088.03
QNG90.3252.0653.1488.2752.0288.83
Q-SAM90.9863.7464.4889.8456.6889.54

The 4–10-qubit simulations apply noise at successive circuit locations, so the reported accuracies reflect accumulated errors.

The theory also accounts for noise accumulation. In the weak local-noise regime, Eq. (20) in Appendix A.1 bounds the accumulated state displacement by ϵn,L=O(nL λ)\epsilon_{n,L}=O(\sqrt{nL}\,\lambda). Under the revised coverage condition in Theorem 4.7, a clean circuit within UG(θ,ϵn,L)U_G(\theta,\epsilon_{n,L}) reproduces the loss-relevant measurement outputs of the noisy state. The maximum loss variation over that region therefore bounds the noisy loss deviation.

We updated Section 5.1 to point directly to Appendix D.2 and Table 13.

Q3. Computational cost as qubit count increases


Appendix B already includes a computational-cost comparison for 4- and 6-qubit circuits (Table 7). We ran this comparison to assess how training cost grows with qubit count. Table 7 reports the number of circuit evaluations per optimizer update for Adam and Q-SAM.

ConfigurationQuantum parametersAdamQ-SAM
4 qubits192384912
6 qubits4328642,040
Increase2.25×2.25×2.24×

Q-SAM’s evaluation count grows at nearly the same rate as Adam’s in this comparison. The block-diagonal QFIM approximation keeps metric construction linear in the parameter count for this architecture, and we amortize that cost by refreshing the metric every 20 updates. The cost of each circuit evaluation also grows with circuit size. For example, dense noisy simulation stores 4n4^n density-matrix entries for an nn-qubit circuit.

We expanded the cost discussion in Section 4.2.2 to point directly to the per-update quantum circuit evaluation counts in Appendix B (Table 7).

Q4. Cross-noise generalization


Our robustness experiments evaluate Q-SAM under noise conditions unseen during training. We train Q-SAM on noise-free circuits and introduce hardware noise only at evaluation. This applies to both the single-source and composite-noise experiments in Section 5.2.

Q-SAM optimizes toward quantum-native flat minima during noise-free training to reduce the sensitivity of the classification loss to hardware noise. In Section 5.3, Q-SAM achieves noisy accuracy close to that of noise-aware training (NAT) under the same composite noise (44.16% vs. 49.42%) and retains higher clean accuracy (76.50% vs. 50.52%). The practical benefit is that Q-SAM achieves this robustness without injecting noise or specifying a device-specific noise model during training.

We believe that this aspect makes Q-SAM naturally become a ‘cross-noise generalizable’ method rather than fitted to a certain type of noise pattern.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer 3eZh,

We appreciate your feedback on the experimental scope and your questions about scaling. With the discussion period ending soon, we wanted to follow up and see if there are any remaining questions or concerns we can address.

Following your suggestions to evaluate Q-SAM beyond image classification and on larger circuits, we added speech and text classification results and experiments with up to 10 qubits.

Reviewer_2XWXOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Q-SAM, which uses quantum information geometry to define the adversarial parameter perturbation in sharpness-aware minimisation. The aim is to improve robustness to quantum hardware noise while training without injected noise. The method is evaluated on image-classification tasks under several noise channels, with additional experiments on IBM hardware.

Critical strengths
  1. Defining sharpness through the quantum-state metric is a well-motivated adaptation of SAM. The comparison with both ordinary SAM and quantum natural gradient helps separate the roles of geometry and the robustness objective.
  2. The empirical results are encouraging across several architectures and noise channels. The hardware experiment provides additional support: on 150 MNIST samples, Q-SAM achieves 49.33% accuracy compared with 27.33% for Adam (Table 3).
Critical weaknesses
  1. The scope of the robustness theorem needs clearer restrictions. Appendix A.3 assumes that the circuit can represent the relevant noise-induced directions, which the local metric alone does not guarantee. Also, a small displacement under a stochastic noise channel does not ensure that every individual Pauli error lies in the same small state-space ball. These conditions should be explicit in Theorem 4.7 rather than inferred from empirical examples.
  2. Practical gains should be assessed against training cost and clean-accuracy tradeoffs. Table 6 reports 912 circuit evaluations per step for Q-SAM versus 384 for Adam in the four-qubit setting. The hardware test is small, and noise-aware training has higher composite-noise accuracy than Q-SAM in Section 5.3, although lower clean accuracy.
Questions for authors
  1. Under what explicit coverage and noise assumptions does Theorem 4.7 hold for the experimental circuits?
  2. Does the robustness advantage persist under matched total training cost, and how variable is the hardware result across repeated runs?
3 comments

We thank the reviewer for the constructive feedback. Your questions helped us clarify the theorem and strengthen our evaluation by measuring robustness and nominal accuracy at matched circuit-evaluation budgets and by quantifying the training cost of noise-aware training.

W1, Q1. Scope and assumptions of Theorem 4.7


We revised Theorem 4.7 to state the measurement coverage and noise assumptions explicitly. When these conditions hold, a parameter perturbation within UG(θ,ϵn,L)U_G(\theta,\epsilon_{n,L}) reproduces the loss-relevant measurement outputs of the noisy state ρ~\tilde{\rho}. Specifically, there exists θ+δ∈UG(θ,ϵn,L)\theta+\delta\in U_G(\theta,\epsilon_{n,L}) satisfying

Lmeas(ρ(θ+δ))=Lmeas(ρ~).L_{\mathrm{meas}}\left(\rho(\theta+\delta)\right)=L_{\mathrm{meas}}(\tilde\rho).

The previous argument bounded the noisy measurement shift by the largest shift over Pauli errors (Appendix A.2). In Appendix A.3 of the original submission, Eq. (38) assumed D(ρ(θ),ρP)≤ϵn,L,ρP=Pρ(θ)P†D(\rho(\theta),\rho_P)\le\epsilon_{n,L},\quad \rho_P=P\rho(\theta)P^\dagger for a fixed Pauli error. Eq. (39) then bounded the corresponding loss change as ∣L(ρP)−L(ρ(θ))∣≤max⁡θ+δ∈UG(θ,ϵn,L)∣L(θ+δ)−L(θ)∣\lvert\mathcal L(\rho_P)-\mathcal L(\rho(\theta))\rvert\le\max_{\theta+\delta\in\mathcal U_G(\theta,\epsilon_{n,L})}\lvert\mathcal L(\theta+\delta)-\mathcal L(\theta)\rvert.

A small displacement of the noisy state does not guarantee Eq. (38) for an individual Pauli error. Even when Eq. (38) holds, the local distance bound alone does not establish that a parameter perturbation can reproduce the relevant change in measured outputs, as assumed in the step to Eq. (39).

The revised proof matches the expectation values used to compute the loss. Under stochastic Pauli noise, these values are the probability-weighted average of the expectations from the Pauli branches. Theorem 4.7 requires the attainable expectations to contain a ball around the nominal expectations and the noisy shift to fit inside that ball.

We also give an explicit condition on the noise strength under which this coverage holds. Appendix A.3 derives it by comparing the size of the shift caused by the channel with the range of shifts the circuit can reproduce within UGU_G.

To derive the coverage radius, we measure parameter changes using the regularized QFIM. Perturbations satisfying ∥(G+κI)1/2δ∥2≤ϵn,L,κ>0\lVert(G+\kappa I)^{1/2}\delta\rVert_2\le\epsilon_{n,L},\quad \kappa>0 lie in the allowed parameter region. We take the Jacobian of the vector of measured expectations with respect to the parameter changes scaled by this metric. We assume it has full row rank at the nominal parameter point. Let σ>0\sigma>0 be its smallest singular value and H>0H>0 a Lipschitz bound on the Jacobian over this region. Then every shift in measured expectations of Euclidean norm at most σr/2\sigma r/2 is attainable, where r=min⁡{ϵn,L,σ/(2H)}r=\min\{\epsilon_{n,L},\sigma/(2H)\}.

The right inverse of the Jacobian gives an initial parameter correction that uses at most half the radius. Within that radius, the Jacobian differs from its nominal value by at most half its smallest singular value. Each subsequent correction is at most half as large as the previous one, so the corrections converge inside the allowed region to a parameter point that reproduces the target expectations.

Let perrp_{\mathrm{err}} be the probability of at least one nonidentity Pauli error. The Euclidean norm of the resulting shift in the measured expectations is at most 2COperr2C_O p_{\mathrm{err}}. The coverage condition therefore holds whenever 2COperr≤σr/2,CO=(∑j∥Oj∥∞2)1/22C_O p_{\mathrm{err}}\leq\sigma r/2,\quad C_O=\bigl(\sum_j\lVert O_j\rVert_\infty^2\bigr)^{1/2}.

For the same input and label, identical measured expectations give identical loss under the same classifier. The noisy loss therefore equals the loss at the matching parameter point.

L(ρ~)=ℓ(Lmeas(ρ~))=ℓ(Lmeas(ρ(θ+δ)))=L(θ+δ).L(\tilde\rho)=\ell\left(L_{\mathrm{meas}}(\tilde\rho)\right)=\ell\left(L_{\mathrm{meas}}(\rho(\theta+\delta))\right)=L(\theta+\delta).

Because the matching parameter point lies in UGU_G, its absolute loss change from the nominal circuit is bounded by the worst-case absolute loss change over that region.

∣L(ρ~)−L(ρ(θ))∣≤max⁡θ+δ′∈UG(θ,ϵn,L)∣L(θ+δ′)−L(θ)∣.\left|L(\tilde\rho)-L(\rho(\theta))\right|\leq\max_{\theta+\delta'\in U_G(\theta,\epsilon_{n,L})}\left|L(\theta+\delta')-L(\theta)\right|.

In Appendix A.3, we replaced the original branchwise argument with this coverage proof and loss bound. The revised argument shows when the channel’s averaged measured expectations can be reproduced by a parameter perturbation within UGU_G.

We updated Lemma 4.6 and stated the coverage and noise conditions directly in Theorem 4.7, using the local-noise radius from Lemma 4.2. Appendix A.3 derives a sufficient coverage radius. Appendix A.4 compares the observed loss changes, and we have not verified the sufficient condition input by input for the experimental circuits.

For the circuits used in our experiments, Theorem 4.7 applies input by input whenever the metric-scaled readout Jacobian has full row rank and the noise-induced shift in measured expectations is at most σr/2\sigma r/2.

W2, Q2. Training cost, practical gains, and hardware variability


We evaluated robustness throughout training by comparing Q-SAM and Adam checkpoints with the same cumulative circuit-evaluation counts. Because Q-SAM requires more circuit evaluations per update, we allowed Adam more updates before each comparison. We averaged the results over five training seeds.

The table below reports the two accuracies at 12 checkpoints. Each cell reports Adam / Q-SAM. Train (Nominal) is the accuracy (%) from the latest completed epoch. Test (Perturbed) is the test accuracy (%) averaged across five noise conditions.

Circuit evaluations (millions)0.070.360.731.462.193.655.116.578.039.4810.5810.80
Updates190 / 80950 / 4001,900 / 8003,800 / 1,6005,700 / 2,4009,500 / 4,00013,300 / 5,60017,100 / 7,20020,900 / 8,80024,700 / 10,40027,550 / 11,60028,120 / 11,840
Train (Nominal)43.7 / 25.865.5 / 54.572.1 / 62.076.3 / 69.578.3 / 72.880.1 / 75.681.3 / 77.881.8 / 78.582.3 / 79.382.6 / 80.082.7 / 80.382.6 / 80.3
Test (Perturbed)34.33 / 22.1738.36 / 40.6835.46 / 42.0837.88 / 41.4239.99 / 44.6836.50 / 48.8035.25 / 51.2234.04 / 51.0634.64 / 53.4435.02 / 54.8834.26 / 54.9135.22 / 53.81

By 6.57 million circuit evaluations, training accuracy was already close to its final value for both optimizers. At that point, Q-SAM achieved 51.06% mean noisy test accuracy vs. 34.04% for Adam. At the final matched checkpoint, Q-SAM’s noisy test accuracy was 15.38–22.64 percentage points higher across the five noise conditions, with similar clean test accuracy (79.30% vs. 80.20%).

The comparison uses the parameter-shift circuit-equivalent counts in Appendix B (Table 7), including Q-SAM’s additional gradient and amortized QFIM evaluations. Table 9 in Appendix B reports both accuracies at all 12 checkpoints.

Q-SAM provides practical robustness gains without explicitly simulating noise during training. In Section 5.3, NAT trains with the same composite noise used for evaluation. Q-SAM approaches NAT’s noisy accuracy without injecting noise during training (44.16% vs. 49.42%) and preserves higher nominal accuracy (76.50% vs. 50.52%).

Under Table 7’s parameter-shift accounting, the dense circuit-simulation costs scale as follows:

CNAT=2P⋅O(B(Ng+NK)4n),CQ-SAM=4P⋅O(BNg2n)+CQFIMR.\begin{aligned} C_{\mathrm{NAT}}&=2P\cdot O\left(B(N_g+N_K)4^n\right),\\ C_{\mathrm{Q\text{-}SAM}}&=4P\cdot O\left(BN_g2^n\right)+\frac{C_{\mathrm{QFIM}}}{R}.\end{aligned}

Here, PP is the number of quantum parameters, BB the batch size, nn the number of qubits, NgN_g the gate count, and NKN_K the additional Kraus-operation count. The final term accounts for QFIM construction amortized over RR updates.

The expressions show why simulating noisy density-matrix evolution increases the work per circuit evaluation. To quantify the actual cost per update, we separately measured full updates in our backpropagation-based simulator using the same four-qubit architecture, batch size 64, and a composite-noise proxy:

MethodSimulationState storageWall-clock time/updateProfiled matrix arithmetic/update
Q-SAMState vectorO(2n)O(2^n)0.54 s29.59 million
NATDensity matrixO(4n)O(4^n)15.13 s1,109.45 million
NAT / Q-SAM——28.0×37.5×

The measurements include forward/backward computation and Q-SAM’s amortized QFIM construction. We report matrix-contraction arithmetic in real-operation equivalents. Section 5.3 reports the simulator measurements, and Appendix B details the circuit-evaluation counts and simulation cost.

These comparisons quantify the practical trade-off. At the same circuit-evaluation budget, Q-SAM improves noisy accuracy over Adam while keeping nominal accuracy close. Compared with NAT, Q-SAM approaches its noisy accuracy, preserves substantially higher nominal accuracy, and reduces the measured simulator update time by about 96%, without requiring a noise model during training.

Hardware variability. Across five independent hardware evaluations on IBM Heron (ibm_torino), Q-SAM achieved (49 ± 1.3)% accuracy and Adam achieved (27 ± 1.7)% accuracy (mean ± sample standard deviation).

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer 2XWX,

Thank you again for your detailed and constructive feedback. With the discussion period ending soon, we wanted to follow up and see if there are any remaining questions or concerns we can address.

Your comments helped us state the assumptions of Theorem 4.7 explicitly and add a robustness comparison at the same total number of training circuit evaluations.

Reviewer_C3hZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Q-SAM, quantum sharpness-aware minimization for training that achieves robustness in Quantum Machine Learning. The main contributions include geometry-based noise reinterpretation and Q-SAM algorithm.

Critical strengths

The trained model was deployed on the physical IBM Heron processor.

Critical weaknesses

The dataset and experiments were done on very small benchmarks.

Questions for authors

How practical is the quantum neural network, cf. the mainstraem NN?

3 comments

We thank the reviewer for the comments. We address the benchmark-size concern and the question of QNN practicality below.

W1. Benchmark size and scope of the experiments


To address the concern about benchmark size, we extended our evaluation to Tiny ImageNet, using the quantum circuit architecture from our CIFAR-10 experiment. Appendix C (Table 10) surveys comparable QNN image classification studies. Prior Tiny ImageNet studies evaluated selected class pairs [1, 2]. Among comparable ICLR, ICML, and NeurIPS main-conference studies we surveyed, the largest label space was the 10 classes of CIFAR-10, evaluated in Schleich et al. (NeurIPS 2024) [3]. Quorus (ICLR 2026) [4] and Wang et al. (ICML 2025) [5] evaluated binary CIFAR-10 tasks. To our knowledge, ours is the first evaluation of a trainable hybrid quantum classifier on all 200 Tiny ImageNet classes under simulated quantum circuit noise.

In the Tiny ImageNet block of Table 2 (details in Appendix D.3, Table 14), all four optimizers achieved similar nominal accuracy (60.72–60.74%). Q-SAM exceeded the strongest baseline, SAM, by 5.67, 3.01, and 5.70 percentage points under depolarizing noise, amplitude damping, and bit flip, respectively. Both phase-noise channels changed accuracy by at most 0.09 percentage points across the four optimizers. This is consistent with the shared Z-basis readout (Section 5.1): phase damping and phase flip preserve Z-basis probabilities when applied immediately before measurement. In-circuit phase errors can still affect the final probabilities through subsequent gates.

OptimizerNominalDepolarizing (0.064)Amplitude damping (0.0775)Phase damping (0.15)Bit flip (0.176)Phase flip (0.05)
Adam60.7332.89 (−27.83)35.44 (−25.29)60.67 (−0.06)33.83 (−26.89)60.64 (−0.09)
SAM60.7436.13 (−24.61)38.04 (−22.70)60.69 (−0.05)37.32 (−23.42)60.69 (−0.05)
QNG60.7428.32 (−32.41)31.76 (−28.98)60.70 (−0.04)29.44 (−31.30)60.71 (−0.02)
Q-SAM60.7241.80 (−18.92)41.04 (−19.68)60.71 (−0.01)43.02 (−17.70)60.71 (−0.01)

The central significance of our experiments lies not in demonstrating robustness on ever-larger benchmarks, but in demonstrating that optimization toward quantum-native flat minima, motivated by our theoretical analysis, improves robustness to hardware noise across diverse image classification benchmarks. We observed this trend on MNIST, Fashion-MNIST, and CIFAR-10, and have now extended it to Tiny ImageNet.

Additional experiments beyond image classification also showed average noisy-accuracy gains over Adam of 7.13 percentage points on Speech Commands and 14.65 percentage points on GoEmotions emotion classification, averaged over five individual noise channels. We report these results in Section 5.2 (Table 4) and provide setup details in Appendix D.4 (Table 15).

Further increases in benchmark size are constrained by the resources available for current quantum experiments. We would therefore appreciate the reviewer’s suggestion of a specific additional benchmark or experimental setting that would most directly address the concern about scale.

Q1. Practicality of QNNs compared with mainstream neural networks


A central motivation for QNN research is their potential for high expressivity with few trainable parameters, supported by studies showing higher effective dimensions than comparable classical neural networks [6]. For example, Senokosov et al. reported comparable or better classification accuracy on MNIST using approximately 12% of the trainable parameters of their classical CNN baseline [7].

Current quantum hardware imposes strict constraints on circuit width and depth, making it infeasible to construct quantum networks at the scale of mainstream classical networks. Under these constraints, improving the reliability of compact QNNs on available hardware is an important step toward practical use.

Hardware noise is a major obstacle to realizing the potential of these models, as performance obtained under ideal conditions can deteriorate substantially during deployment. Q-SAM directly addresses this obstacle by optimizing toward quantum-native flat minima, improving hardware-noise robustness without requiring noise injection or a device-specific noise model during training. In our IBM Heron experiment, Q-SAM achieved 49% accuracy vs. 27% for Adam, an improvement of 22 percentage points using the same circuit architecture.

Our comparison with noise-aware training (NAT) further illustrates this practical benefit. Q-SAM achieved 44% composite-noise accuracy vs. 49% for NAT, while preserving substantially higher clean accuracy (77% vs. 51%). In a separate simulator timing comparison using the same four-qubit architecture, Q-SAM needed only 0.54 seconds per training update vs. 15.13 seconds for NAT. NAT updates took approximately 28 times longer, while Q-SAM reduced the time per update by 96% (detailed in our response to Reviewer 2XWX). Q-SAM therefore approaches NAT’s noisy accuracy while preserving clean performance and avoiding explicit noise simulation during training. Together with the robustness improvements across the classification benchmarks, these results demonstrate our contribution to QNN practicality.

If the reviewer has a specific criterion of practicality in mind, we would appreciate clarification so that we can address the question more directly.

References

[1] K. Phalak, A. Ghosh, and S. Ghosh. “Optimizing Quantum Embedding using Genetic Algorithm for QML Applications.” ISQED, 2025.

[2] Y. Kim et al. “Multi-Channel Convolutional Neural Quantum Embedding.” Advanced Quantum Technologies, 2026.

[3] P. Schleich et al. “Quantum Deep Equilibrium Models.” NeurIPS, 2024.

[4] J. Han et al. “Layerwise Federated Learning for Heterogeneous Quantum Clients using Quorus.” ICLR, 2026.

[5] X. Wang, H. Tao, and R. Wu. “Predictive Performance of Deep Quantum Data Re-uploading Models.” ICML, 2025.

[6] A. Abbas et al. “The power of quantum neural networks.” Nature Computational Science, 2021.

[7] A. Senokosov et al. “Quantum machine learning for image classification.” Machine Learning: Science and Technology, 2024.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer C3hZ,

Thank you for taking the time to review our paper. With the discussion period ending soon, we wanted to follow up and see if there are any remaining questions or concerns we can address.

In response to your questions about benchmark size and the practicality of QNNs, we added Tiny ImageNet results and clarified the scope of our robustness claims.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Summary of manuscript revisions

Dear Reviewers,

Thank you for your constructive feedback. Below, we summarize the main revisions we made in response to your comments. Changes are highlighted in blue\color{blue}{\mathrm{blue}} in the manuscript.

We expanded Q-SAM’s evaluation to a larger image benchmark (Tiny ImageNet) and to tasks beyond image classification (speech and text classification). We also added training cost comparisons to quantify the practical benefits of Q-SAM and clarified the conditions of the theoretical robustness bound.

1. Evaluation on a larger image classification benchmark (Reviewer C3hZ)

  • We added robustness results on all 200 classes of Tiny ImageNet (Section 5.2, Table 2). Under depolarizing noise, amplitude damping, and bit flips, Q-SAM improved test accuracy by 5.6–9.2 percentage points over Adam and 3.0–5.7 points over the strongest baseline, with nearly identical nominal accuracy across optimizers.

2. Robustness beyond image classification (Reviewers 3eZh and C3hZ)

  • We added robustness experiments on spoken command recognition and text emotion classification (Section 5.2, Table 4). Q-SAM achieved the highest test accuracy under each of the five noise channels on both tasks. Averaged across channels, Q-SAM improved test accuracy by 7.13 points over Adam and 3.31 points over the strongest baseline on Speech Commands, and by 14.65 and 7.62 points, respectively, on GoEmotions.

3. Practical benefits of Q-SAM and a fair robustness comparison (Reviewers 2XWX and C3hZ)

  • We revised Section 5.3 to quantify the computational cost of noise-aware training (NAT) relative to Q-SAM’s standard noise-free training. In our simulator, the matrix calculations in a NAT update cost 37.5 times as much as those in a Q-SAM update, while the previously reported composite-noise test accuracy of Q-SAM was only 5.26 percentage points below NAT’s (44.16% vs. 49.42%).
  • Given that Q-SAM requires more circuit evaluations per update than Adam, we added a robustness comparison that matches the total number of training circuit evaluations at several stages of training (Appendix B, Table 9). From roughly the midpoint of training onward, Q-SAM achieved 16–21 percentage points higher average test accuracy under noise than Adam at the same circuit evaluation count.

4. Clarification of the theoretical robustness bound (Reviewer 2XWX)

  • We clarified the assumptions behind Theorem 4.7. Appendix A.3 now gives an explicit sufficient condition on the noise strength. Under that condition and the stated circuit assumptions, an allowed change to the circuit parameters can reproduce the measured outputs used to compute the noisy loss. The noisy loss therefore equals the loss at one allowed parameter perturbation, so the worst-case loss change over the allowed perturbations bounds the loss change caused by noise.
AuthorsOpenReview ↗

Follow-up on our responses

Dear Reviewers,

With about 24 hours remaining in the discussion period, we would appreciate any follow-up on our responses. If any concerns remain unresolved or need further clarification, please let us know so that we can address them before the discussion closes.

Thank you for your time.

Dear Reviewers,

Although the discussion period is nearing its end, if you have any further comments, we will do our best to address them. Even a brief response would be greatly appreciated. We remain available to respond promptly until the discussion closes.

Thank you again for your time and consideration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.