Pith. sign in

REVIEW 4 major objections 5 minor 48 references

DistributedEstimator: Distributed Training of Quantum Neural Networks via Circuit Cutting

T0 review · 4 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Circuit cutting for quantum neural network training is bottlenecked by classical reconstruction, not quantum execution—a serial step that eats 53% of per-query time at three cuts and caps parallel speed-up.

desk verdict First measured stage-level timing of circuit cutting in a QNN training loop puts reconstruction at a 43–53% share on a simulator — a useful methodology and a useful warning, but the Amdahl bound is not backend-independent. read the letter →

arxiv 2602.16233 v3 pith:UHVETEIQ submitted 2026-02-18 cs.DC cs.LGquant-ph

Prabhjot Singh , Adel N. Toosi , Rajkumar Buyya This is my paper
classification cs.DCcs.LGquant-ph
keywords circuitcuttingquantumneuralnetworksdistributedtrainingreconstructionbottleneckAmdahl'slawmachinelearningvariationalalgorithmsrobustness
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Circuit cutting promises to run quantum neural networks that exceed device size, but this paper argues that its true cost is a systems problem: each training query becomes a staged distributed workload whose final classical reconstruction is the bottleneck. By instrumenting every estimator query across partitioning, subexperiment generation, parallel execution, and reconstruction, the authors measure a median reconstruction share of 53% (95th percentile 58%) of per-query time at three cuts, which by Amdahl's law bounds achievable speed-up near 1.9x regardless of worker count. Training time grows multiplicatively because each query expands into O(9^c) subexperiments, and parameter-shift gradients multiply that expansion by additional queries per parameter. A further finding is outcome preservation: test accuracy is fully preserved on Iris and not systematically degraded on MNIST, and robustness to Gaussian and gradient-based perturbations is maintained or occasionally improved. The paper therefore establishes that scaling cut-based QNN training requires treating reconstruction as a first-class parallel or approximation target, not just adding workers.

What carries the argument

The central object is the four-stage estimator pipeline—partitioning, subexperiment generation, parallel execution, and classical reconstruction—with the timing decomposition T_total = T_part + T_gen + T_exec + T_rec. Reconstruction acts as a global barrier: every subexperiment result must arrive before the estimate is formed. The load-bearing quantity is the serial fraction f = T_rec/T_total; Amdahl's law applied to this pipeline gives maximum speed-up 1/(f + (1-f)/w), which converges to 1/f, so the measured f=0.53 caps speed-up near 1.9x. A second mechanism is exponential subexperiment expansion O(9^c) from CNOT-based wire cutting, compounded per estimator query by the parameter-shift grad

What would settle it

Measure T_rec/T_total and speed-up at 16 workers for the same three-cut configuration on a process- or node-level distributed runtime, or with subexperiment service times artificially varied across backends. If the median reconstruction share falls well below 53% once the worker pool is isolated, or if speed-up grows beyond the 1/f Amdahl bound, the reconstruction-dominance claim is a harness artefact rather than a pipeline property.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that classical reconstruction, not quantum subcircuit execution, dictates wall-clock time in cut-based QNN training. Per estimator query, reconstruction rises from a median 43% of total time at one cut to 53% at three cuts, making the pipeline reduction-bound; because reconstruction is serial and waits on all results, additional workers do not translate into proportional speed-up. At the same time, cutting preserves the trained model: Iris accuracy is identical across all cut settings, MNIST shows no systematic degradation, and robustness under input noise and adversarial perturbations is maintained, with some cut configurations matching or

Load-bearing premise

The load-bearing premise, asserted in Section 6.2, is that the measured stage shares and scaling trends are structural properties of the cut pipeline rather than artefacts of the test harness—specifically, a thread-based Python worker pool driving a local simulator, with no control experiment separating thread-pool overhead or simulator internals from pipeline-stage cost.

Editorial extensions

If this is right

  • At three cuts, the median reconstruction share of 53% implies a theoretical maximum speed-up of about 1.9x from unlimited parallelism; no number of workers can do better without shrinking f.
  • Training time grows multiplicatively as O(9^c) per estimator query, and parameter-shift gradients double the query count per parameter, so an order-of-magnitude training time increase at three cuts is a direct consequence.
  • In reconstruction-dominated configurations, execution-side stragglers matter less: injected delays produce smaller relative slowdowns, so speculative execution or work stealing yields diminishing returns unless the reconstruction phase itself is improved.
  • Cutting preserves learning outcomes: accuracy on Iris is identical across no-cut through three-cut settings, MNIST accuracy shows no systematic degradation, and robustness to Gaussian noise and gradient-based attacks is maintained or occasionally improved.
  • The O(9^c) expansion bounds practical experimentation to small qubit counts with current cutting schemes, reinforcing that approximate or distributed reconstruction is a prerequisite for scaling.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the reconstruction share holds at larger scales, the most direct fix is to overlap reconstruction with late subexperiment arrivals—incremental aggregation would convert serial reconstruction time into pipeline latency and effectively lower the Amdahl fraction; this is a design the paper proposes but does not test.
  • The occasional robustness gains at higher cut counts suggest cut-induced estimator noise during training may act as an implicit regulariser; the paper notes this possibility, but its limited seed count does not establish it.
  • Because the bottleneck is classical, a natural benchmark is to compare cut-aware training against an equivalent classical-width model or against cheaper approximate reconstruction; such a comparison would show whether cutting still wins when reconstruction is optimised.
  • The 9-fold base of O(9^c) is not fundamental—randomised or communication-assisted cutting reduces it—so the paper's scale limit is specific to CNOT-based wire cutting rather than to circuit cutting in general; swapping the decomposition scheme should lower the reconstruction share and extend the practical range.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes DistributedEstimator, an instrumented estimator-execution pipeline for training quantum neural networks with circuit cutting. Each estimator query is decomposed into four measured stages — partitioning, subexperiment generation, parallel execution, and classical reconstruction — and evaluated on Iris and MNIST binary classification tasks using Qiskit/Aer and a thread-based worker pool. The central empirical claim is that classical reconstruction dominates per-query time (median T_rec/T_total = 0.53, p95 = 0.58 at three cuts), so Amdahl's law bounds achievable speed-up to roughly 1.9x regardless of worker count, while accuracy and robustness are preserved relative to the uncut baseline. The paper also studies straggler sensitivity and argues that reconstruction-dominated regimes are less sensitive to execution-side stragglers. The authors position these findings as structural properties of the cut-aware pipeline rather than artefacts of the specific runtime.

Significance. If the measured reconstruction-dominance and scaling saturation generalize beyond the Aer-simulator setup, the paper provides a useful systems-level characterization of circuit cutting for QML training: it identifies the reconstruction barrier as a first-class optimization target and gives concrete implications for scheduling and resource allocation. The paper ships a reproducible instrumentation methodology, logs stage-level timings, and explicitly discloses its limitations (e.g., n=5 for the 3-cut setting, thread-based pool, small qubit counts). These strengths are real: the matched-pair scaling ratios and the transparent treatment of Eq. (1) as a measured decomposition are methodologically sound. However, the central quantitative claim is measured only on a fast simulator, and the paper asserts backend-independence without comparative evidence; this external-validity risk is the main factor preventing acceptance in current form.

major comments (4)
  1. [§5.2.2/Table 1] The reconstruction share T_rec/T_total is central to the paper, but Table 1 does not state the worker pool size used when computing T_rec/T_total. Since T_rec is serial and T_exec shrinks with more workers, the ratio is worker-dependent; a median of 0.53 at, say, w=16 would be materially different at w=1. Without specifying w, the Amdahl bound of ~1.9x is not reproducible. Please report the worker count for each row, or provide the full distribution of T_rec/T_total across worker counts.
  2. [§5.1 and §6.2] All measurements use an Aer simulator with per-subexperiment execution times in milliseconds. The manuscript claims in §6.2 that the findings are 'structural properties of the pipeline rather than threading artefacts', but no evidence supports backend-independence. On a real QPU, per-circuit latency includes queueing, transmission, and shot accumulation, which would inflate T_exec and reduce T_rec/T_total; the observed 53% reconstruction share may therefore be an artefact of the simulator's fast execution. As this is load-bearing for the Amdahl-limited speed-up claim, the authors should either (a) validate on a remote/hardware backend, (b) provide a latency-model sensitivity analysis showing how T_rec/T_total varies with realistic T_exec values, or (c) explicitly restrict the claim to simulator-class backends.
  3. [§5.2.2/Table 1, §4.1] The headline statistics for the three-cut configuration rest on n=5 runs (Table 1), and no error bars or confidence intervals appear in Figures 4–8 or Table 1. The monotonic increase in reconstruction share across cut counts is directionally suggestive, but with n=5 and no variance measure, the specific median 0.53 and p95 0.58 are not statistically anchored. Please provide per-run values, bootstrap intervals, or at least the number of seeds for each bar in the figures, and explicitly discuss the uncertainty around the 53% claim.
  4. [§4.2 and §5.2.3] RQ3 is framed as studying scheduling policies that stagger or reshape dispatch, and 'scheduling policy' is listed as a contribution (§1.4). However, the evaluation only exercises eager dispatch with synthetic straggler injection (Fig. 6); batched and staggered dispatch strategies are described in §4.2 but not measured. The paper's future-work paragraph correctly acknowledges this gap, but then §5.2.3 and §6.1 draw conclusions about straggler mitigation strategies that are not tested. Either add experiments comparing eager vs. batched vs. staggered policies, or revise the RQ3 framing and contribution claims to state that only straggler sensitivity under eager dispatch is measured.
minor comments (5)
  1. [§5.2.1] The statement that MNIST training time increases 'approximately an order of magnitude' from no cut to three cuts would be more useful with the actual measured times or a range across seeds, rather than an informal approximation.
  2. [§4.2] Algorithm 2 calls π an 'ordering rule' but the three policies in the text differ in batch size B, inter-batch delay δ, and grouping. Please align the notation so that π is defined consistently as a tuple (B, δ, ordering) and state which values are used for the eager baseline.
  3. [§5.2.4] The explanation that cutting 'can occasionally lead to improved final accuracy by effectively regularising the optimisation through increased estimator noise' is plausible but speculative; since RQ4 only requires absence of systematic degradation, the sentence could be softened and left to future work.
  4. [§6.2] The phrase 'structural properties of the pipeline' is repeated from the abstract and conclusions; given the simulator-only evidence, this wording overstates generality. A more modest formulation (e.g., 'observed in this simulator-based evaluation') would better match the evidence.
  5. [§5.1] The text says 'Iris: maxiter=60' in one place and 'Iris: maxiter=10' for scaling runs; consider adding a table of experimental configurations to make the budget matching unambiguous.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: all load-bearing claims are empirical measurements or externally imported tooling properties, with no fitted parameter or self-citation chain used as the derivation.

full rationale

The paper is an empirical measurement study, and its derivation chain is not circular. The central claims—reconstruction share T_rec/T_total of 53% median at three cuts, speed-up saturation near 1, and preserved accuracy/robustness—are directly measured from instrumented runtime logs and training traces (Sections 5.1–5.2). Equation (1) is an additive stage decomposition by definition, and Eq. (2) is a modelling approximation of parallel execution; neither is used to infer the measured quantities. The Amdahl-style bound 1/(f + (1-f)/w) is applied post hoc to the measured serial fraction f, not used to fit or construct f, so it is arithmetic applied to data rather than a prediction that reduces to its inputs. The O(9^c) subexperiment growth is imported from qiskit-addon-cutting [18] and CutQC [10], which are external tooling/literature, not self-citations. The only self-referential item is the closing pointer to QFaaS [48] in Section 7, which is a future-work remark and is not load-bearing for any result. Section 6.2's assertion that findings are 'structural properties of the pipeline rather than threading artefacts' is a generalisability/limitation claim lacking comparative evidence, but lack of external validation is a correctness-risk issue, not circularity. No parameter is fitted and then renamed a prediction, no uniqueness theorem is imported from the authors' own work, and no ansatz is smuggled in via self-citation.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper is an empirical systems study: it introduces no new formal quantities. Its claims rest on standard QML/cutting assumptions and on hand-chosen experimental hyperparameters (shot budget, straggler delay/probability). The main epistemic risk is the proxy assumptions — Aer simulator plus thread pool — rather than hidden fitted parameters.

free parameters (3)
  • per-subexperiment shot budget S = 1024
    Fixed across baseline and cut configurations (Sec. 5.1). The accuracy and robustness preservation claims are conditional on this finite-sample noise level, and no sensitivity sweep over shots is reported.
  • straggler_delay_s = 0.1
    Chosen for the RQ3 straggler injection experiments (Sec. 5.1). The qualitative conclusion that reconstruction-dominated configurations are less sensitive to execution-side stragglers is tested at a single delay magnitude relative to simulator subexperiment times.
  • straggler_probability_p = 0.2
    Chosen for RQ3 (Sec. 5.1). Only one probability level is evaluated, so the straggler-sensitivity ordering is established at a single operating point.
assumptions (4)
  • domain assumption Circuit cutting reconstruction is unbiased: reconstructed expectation values equal the original to within finite-sample variance.
    Used throughout Sec. 3.2 and 5.2.4 to argue accuracy should be preserved; the paper relies on the qiskit-addon-cutting implementation rather than proving this.
  • domain assumption Parameter-shift gradient queries are mathematically equivalent with and without cutting — each shifted query is independently cut and reconstructed.
    Sec. 3.2 justifies the claim that optimizer behavior differs only via estimator variance, not via algorithmic change.
  • domain assumption A thread-based worker pool execution on an Aer simulator is representative of distributed subexperiment execution for scaling and straggler conclusions.
    Secs. 5.1 and 6.2. The paper asserts findings are structural properties rather than threading artifacts, without a comparative distributed runtime.
  • domain assumption Aer simulator execution time is a valid proxy for subexperiment service times when computing reconstruction share and speed-up.
    Sec. 5.1. On real hardware the T_exec share would change; the 53% reconstruction figure is specific to simulator-scale latencies.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DistributedEstimator: Distributed Training of Quantum Neural Networks via Circuit Cutting." pith.science (2026). https://pith.science/paper/UHVETEIQ

@misc{pith2026260216233,
  author       = {Pith},
  title        = {Pith review of: DistributedEstimator: Distributed Training of Quantum Neural Networks via Circuit Cutting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UHVETEIQ}},
  note         = {Machine review of arXiv:2602.16233}
}
abstract

Circuit cutting decomposes a large quantum circuit into smaller subcircuits executed independently; expectation values are recovered by classically combining subcircuit outcomes. Prior work characterises cutting overhead via subcircuit counts and sampling complexity, but its end-to-end impact on iterative, estimator-driven training pipelines remains under-measured from a systems perspective. We propose DistributedEstimator, a cut-aware estimator execution pipeline that treats circuit cutting as a staged distributed workload, instrumenting each query across four phases: partitioning, subexperiment generation, parallel execution, and classical reconstruction. Using logged runtime traces and learning outcomes on two binary classification workloads (Iris and MNIST), we quantify cutting overheads, scaling limits, and sensitivity to injected stragglers, and assess whether accuracy and robustness are preserved under matched training budgets. Reconstruction dominates per-query time -- a median of 53% and 95th percentile of 58% at three cuts -- bounding achievable speed-up under parallelism. Despite this, test accuracy is fully preserved on Iris and maintained without systematic degradation on MNIST across all cut configurations. Robustness under Gaussian noise and FGSM perturbations is similarly preserved, with several configurations matching or improving on the uncut baseline. Exponential growth of subexperiment counts (${O}(9^c)$ for CNOT-based decomposition) is a fundamental barrier limiting practical experimentation to small qubit counts. These results establish that practical scaling for learning workloads requires reducing and overlapping reconstruction, scheduling policies for barrier-dominated critical paths, and computationally efficient reconstruction strategies for larger qubit counts.

Figures

Figures reproduced from arXiv: 2602.16233 by the authors.

Figure 1
Figure 1. Independent subcircuits generated by cutting a circuit. The [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Training pipeline with a cut-aware distributed estima [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Per-query expansion under circuit cutting. Subexperiments [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: RQ2: Scaling behaviour under clean execution. Legend: [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 7
Figure 7. Figure 7: RQ4: Absolute test accuracy under clean execution. Leg [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: RQ5: Robustness summary under clean execution. Legend: [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 8 linked inside Pith

  1. [1]

    Preskill, Quantum computing in the NISQ era and beyond, Quantum 2 (2018) 79.doi:10

    J. Preskill, Quantum computing in the NISQ era and beyond, Quantum 2 (2018) 79.doi:10. 22331/q-2018-08-06-79

    2018

  2. [2]

    Cerezo, A

    M. Cerezo, A. Arrasmith, R. Babbush, S. C. Ben- jamin, S. Endo, K. Fujii, J. R. McClean, K. Mi- tarai, X. Yuan, L. Cincio, P. J. Coles, Variational quantum algorithms, Nature Reviews Physics 3 (9) (2021) 625–644.doi:10.1038/s42254-021- 00348-9

    doi:10.1038/s42254-021- 2021

  3. [3]

    S. Endo, Z. Cai, S. C. Benjamin, X. Yuan, Hy- brid quantum-classical algorithms and quantum error mitigation, Journal of the Physical Society of Japan 90 (3) (2021) 032001.doi:10.7566/ JPSJ.90.032001

    2021

  4. [4]

    Schuld, V

    M. Schuld, V . Bergholm, C. Gogolin, J. Izaac, N. Killoran, Evaluating analytic gradients on quantum hardware, Physical Review A 99 (3) (2019) 032331.doi:10.1103/PhysRevA.99. 032331

    doi:10.1103/physreva.99 2019

  5. [5]

    Temme, S

    K. Temme, S. Bravyi, J. M. Gambetta, Error miti- gation for short-depth quantum circuits, Physical Review Letters 119 (18) (2017) 180509.doi: 10.1103/PhysRevLett.119.180509

    doi:10.1103/physrevlett.119.180509 2017

  6. [6]

    J. Dean, S. Ghemawat, MapReduce: Simplified data processing on large clusters, in: Proceedings of the 6th USENIX Symposium on Operating Systems Design and Implementation (OSDI), 2004, pp. 137–150. URLhttps://research.google.com/ archive/mapreduce-osdi04.pdf

    2004

  7. [7]

    Zaharia, A

    M. Zaharia, A. Konwinski, A. D. Joseph, R. Katz, I. Stoica, Improving MapReduce performance in heterogeneous environments, in: Proceedings of the 8th USENIX Symposium on Operating Systems Design and Implementation (OSDI), 2008. URLhttps://www.usenix.org/event/ osdi08/tech/full_papers/zaharia/

    2008

  8. [8]

    Sergeev, M

    A. Sergeev, M. Del Balso, Horovod: fast and easy distributed deep learning in TensorFlow, arXiv preprint arXiv:1802.05799 (2018).doi: 10.48550/arXiv.1802.05799

    arXiv doi:10.48550/arxiv.1802.05799 2018

Show all 48 references
  1. [9]

    T. Peng, A. W. Harrow, M. Ozols, X. Wu, Sim- ulating large quantum circuits on a small quan- tum computer, Physical Review Letters 125 (15) (2020) 150504.doi:10.1103/PhysRevLett. 125.150504

    2020 doi

  2. [10]

    W. Tang, T. Tomesh, M. Suchara, J. Larson, M. Martonosi, CutQC: Using small quantum com- puters for large quantum circuit evaluations, in: Proceedings of the 26th ACM International Con- ference on Architectural Support for Programming Languages and Operating Systems (ASPLOS), ...

    2021 doi

  3. [11]

    Piveteau, D

    C. Piveteau, D. Sutter, Circuit knitting with clas- sical communication, IEEE Transactions on Infor- mation Theory 70 (5) (2024) 3001–3016, preprint: arXiv:2205.00016.doi:10.1109/TIT.2023. 3310797

    2024 arXiv

  4. [12]

    A. W. Harrow, A. Lowe, Optimal quantum cir- cuit cuts with application to clustered Hamiltonian simulation, PRX Quantum 6 (1) (2025) 010316. doi:10.1103/PRXQuantum.6.010316

    2025 doi

  5. [13]

    A. Lowe, M. Medvidovi ´c, A. Hayes, L. J. Romero, N. M. Tubman, D. Camps, Fast quantum circuit 15 cutting with randomized measurements, Quantum 7 (2023) 934.doi:10.22331/q-2023-03-02- 934

    2023 doi

  6. [14]

    Madry, A

    A. Madry, A. Makelov, L. Schmidt, D. Tsipras, A. Vladu, Towards deep learning models resistant to adversarial attacks, in: International Confer- ence on Learning Representations (ICLR), 2018. URLhttps://openreview.net/forum?id= rJzIBfZAb

    2018

  7. [15]

    Tramèr, A

    F. Tramèr, A. Kurakin, N. Papernot, I. Goodfel- low, D. Boneh, P. McDaniel, Ensemble adversarial training: Attacks and defenses, in: International Conference on Learning Representations (ICLR), 2018. URLhttps://openreview.net/forum?id= rkZvSe-RZ

    2018

  8. [16]

    Moritz, R

    P. Moritz, R. Nishihara, S. Wang, A. Tumanov, R. Liaw, E. Liang, M. Elibol, Z. Yang, W. Paul, M. I. Jordan, I. Stoica, Ray: A distributed framework for emerging AI applications, in: Proceedings of the 13th USENIX Symposium on Operating Systems Design and Implementation (OSDI),...

    2018

  9. [17]

    Qiskit Machine Learning Contributors, Qiskit machine learning,https://qiskit.org/ ecosystem/machine-learning/, accessed: 2026-04 (2024)

    2026

  10. [18]

    io/qiskit-addon-cutting/, accessed: 2026- 04 (2024)

    Qiskit Addon Cutting Contributors, Circuit cutting with Qiskit addons,https://qiskit.github. io/qiskit-addon-cutting/, accessed: 2026- 04 (2024)

    2026

  11. [19]

    Peruzzo, J

    A. Peruzzo, J. McClean, P. Shadbolt, M.-H. Yung, X.-Q. Zhou, P. J. Love, A. Aspuru-Guzik, J. L. O’Brien, A variational eigenvalue solver on a pho- tonic quantum processor, Nature Communications 5 (2014) 4213.doi:10.1038/ncomms5213

    2014 doi

  12. [20]

    N. Moll, P. Barkoutsos, L. S. Bishop, J. M. Chow, A. Cross, D. J. Egger, S. Filipp, A. Fuhrer, J. M. Gambetta, M. Ganzhorn, A. Kandala, et al., Quan- tum optimization using variational algorithms on near-term quantum devices, Quantum Science and Technology 3 (3) (2018) 030503....

    2018

  13. [21]

    Kandala, A

    A. Kandala, A. Mezzacapo, K. Temme, M. Takita, M. Brink, J. M. Chow, J. M. Gambetta, Hardware- efficient variational quantum eigensolver for small molecules and quantum magnets, Nature 549 (7671) (2017) 242–246.doi:10.1038/ nature23879

    2017

  14. [22]

    Biamonte, P

    J. Biamonte, P. Wittek, N. Pancotti, P. Rebentrost, N. Wiebe, S. Lloyd, Quantum machine learning, Nature 549 (7671) (2017) 195–202.doi:10. 1038/nature23474

    2017

  15. [23]

    Mitarai, M

    K. Mitarai, M. Negoro, M. Kitagawa, K. Fu- jii, Quantum circuit learning, Physical Review A 98 (3) (2018) 032309.doi:10.1103/PhysRevA. 98.032309

    2018 doi

  16. [24]

    Havlíˇcek, A

    V . Havlíˇcek, A. D. Córcoles, K. Temme, A. W. Harrow, A. Kandala, J. M. Chow, J. M. Gambetta, Supervised learning with quantum-enhanced fea- ture spaces, Nature 567 (7747) (2019) 209–212. doi:10.1038/s41586-019-0980-2

    2019 doi

  17. [25]

    Huang, M

    H.-Y . Huang, M. Broughton, M. Mohseni, R. Bab- bush, S. Boixo, H. Neven, J. R. McClean, Power of data in quantum machine learning, Nature Com- munications 12 (2021) 2631.doi:10.1038/ s41467-021-22539-9

    2021

  18. [26]

    J. R. McClean, S. Boixo, V . N. Smelyanskiy, R. Babbush, H. Neven, Barren plateaus in quan- tum neural network training landscapes, Nature Communications 9 (2018) 4812.doi:10.1038/ s41467-018-07090-4

    2018

  19. [27]

    Javadi-Abhari, M

    A. Javadi-Abhari, M. Treinish, K. Krsulich, C. J. Wood, J. Lishman, J. Gacon, S. Martiel, P. D. Nation, L. S. Bishop, A. W. Cross, B. R. John- son, J. M. Gambetta, Quantum computing with Qiskit, arXiv preprint arXiv:2405.08810 (2024). doi:10.48550/arXiv.2405.08810

    2024 doi arXiv

  20. [28]

    URLhttps://quantum.cloud.ibm.com/ docs/api/qiskit/primitives

    IBM Quantum, Qiskit primitives API docu- mentation (Estimator and Sampler), Online documentation, accessed: 2026-04 (2024). URLhttps://quantum.cloud.ibm.com/ docs/api/qiskit/primitives

    2026

  21. [29]

    Bergholm, J

    V . Bergholm, J. Izaac, M. Schuld, C. Gogolin, N. Killoran, et al., PennyLane: Automatic dif- ferentiation of hybrid quantum-classical compu- tations, arXiv preprint arXiv:1811.04968 (2018). doi:10.48550/arXiv.1811.04968. 16

    2018 doi arXiv

  22. [30]

    Broughton, G

    M. Broughton, G. Verdon, T. McCourt, A. J. Martinez, J. H. Yoo, S. V . Isakov, P. Massey, R. Halavati, M. Y . Niu, A. Zlokapa, et al., TensorFlow Quantum: A software framework for quantum machine learning, arXiv preprint arXiv:2003.02989 (2020).doi:10.48550/ arXiv.2003.02989

    2003 doi arXiv

  23. [31]

    Zaharia, M

    M. Zaharia, M. Chowdhury, T. Das, A. Dave, J. Ma, M. McCauley, M. J. Franklin, S. Shenker, I. Stoica, Resilient distributed datasets: A fault-tolerant abstraction for in-memory cluster computing, in: Proceedings of the 9th USENIX Symposium on Networked Systems Design and Imple...

    2012

  24. [32]

    Rocklin, Dask: Parallel computation with blocked algorithms and task scheduling, in: Pro- ceedings of the 14th Python in Science Conference (SciPy), 2015, pp

    M. Rocklin, Dask: Parallel computation with blocked algorithms and task scheduling, in: Pro- ceedings of the 14th Python in Science Conference (SciPy), 2015, pp. 126–132.doi:10.25080/ Majora-7b98e3ed-013

    2015

  25. [33]

    M. Li, D. G. Andersen, J. W. Park, A. J. Smola, A. Ahmed, V . Josifovski, J. Long, E. J. Shekita, B.-Y . Su, Scaling distributed machine learning with the parameter server, in: Proceedings of the 12th USENIX Symposium on Operating Systems Design and Implementation (OSDI), 2014...

    2014

  26. [34]

    Q. Ho, J. Cipar, H. Cui, J. K. Kim, S. Lee, P. B. Gibbons, G. Gibson, G. R. Ganger, E. P. Xing, More effective distributed ML via a stale synchronous parallel parameter server, in: Ad- vances in Neural Information Processing Systems (NeurIPS), 2013. URLhttps://papers.neurips.c...

    2013

  27. [35]

    Recht, C

    B. Recht, C. Re, S. Wright, F. Niu, Hogwild!: A lock-free approach to parallelizing stochastic gra- dient descent, in: Advances in Neural Information Processing Systems (NeurIPS), 2011. URLhttps://proceedings. neurips.cc/paper/2011/hash/ 218a0aefd1d1a4be65601cc6ddc1520e

    2011

  28. [36]

    Cho, et al., BlueConnect: Decomposing all-reduce for deep learning on heterogeneous network hierarchy, in: Proceedings of Machine Learning and Systems (MLSys), 2019

    M. Cho, et al., BlueConnect: Decomposing all-reduce for deep learning on heterogeneous network hierarchy, in: Proceedings of Machine Learning and Systems (MLSys), 2019. URLhttps://mlsys.org/Conferences/ 2019/doc/2019/130.pdf

    2019

  29. [37]

    Goyal, P

    P. Goyal, P. Dollár, R. B. Girshick, P. No- ordhuis, L. Wesolowski, A. Kyrola, A. Tul- loch, Y . Jia, K. He, Accurate, large mini- batch SGD: Training ImageNet in 1 hour, arXiv preprint arXiv:1706.02677 (2017).doi:10. 48550/arXiv.1706.02677

    2017 doi arXiv

  30. [38]

    R. D. Blumofe, C. E. Leiserson, Scheduling mul- tithreaded computations by work stealing, Jour- nal of the ACM 46 (5) (1999) 720–748.doi: 10.1145/324133.324234

    1999

  31. [39]

    Carrera Vazquez, C

    A. Carrera Vazquez, C. Tornow, D. Ristè, S. Wo- erner, M. Takita, D. J. Egger, Combining quan- tum processors with real-time classical communi- cation, Nature 636 (2024) 75–79.doi:10.1038/ s41586-024-08178-2

    2024

  32. [40]

    Garrison, et al., Circuit cutting with Quantum Serverless, IBM Research, accessed: 2026-04 (2023)

    J. Garrison, et al., Circuit cutting with Quantum Serverless, IBM Research, accessed: 2026-04 (2023). URLhttps://research.ibm.com/ publications/circuit-cutting-with- quantum-serverless

    2026

  33. [41]

    Sitdikov, et al., Circuit knitting toolbox and quantum serverless, IBM Research, accessed: 2026-04 (2023)

    I. Sitdikov, et al., Circuit knitting toolbox and quantum serverless, IBM Research, accessed: 2026-04 (2023). URLhttps://research.ibm.com/ publications/circuit-knitting- toolbox-and-quantum-serverless

    2026

  34. [42]

    Johnson, et al., Simulating larger quantum circuits with circuit cutting and Quantum Server- less, SC 2023 Poster, IBM Research, accessed: 2026-04 (2023)

    C. Johnson, et al., Simulating larger quantum circuits with circuit cutting and Quantum Server- less, SC 2023 Poster, IBM Research, accessed: 2026-04 (2023). URLhttps://research.ibm.com/ publications/simulating-larger- quantum-circuits-with-circuit- cutting-and-quantum-serverless

    2023

  35. [43]

    Szegedy, W

    C. Szegedy, W. Zaremba, I. Sutskever, J. Bruna, D. Erhan, I. Goodfellow, R. Fergus, In- triguing properties of neural networks, arXiv preprint arXiv:1312.6199 (2013).doi:10. 48550/arXiv.1312.6199

    2013 doi arXiv

  36. [44]

    I. J. Goodfellow, J. Shlens, C. Szegedy, Explain- ing and harnessing adversarial examples, in: Inter- national Conference on Learning Representations 17 (ICLR), 2015. URLhttps://arxiv.org/abs/1412.6572

    2015 arXiv

  37. [45]

    Qiskit Contributors, Qiskit: An open-source framework for quantum computing,https:// qiskit.org, accessed: 2026-04 (2024)

    2026

  38. [46]

    LeCun, L

    Y . LeCun, L. Bottou, Y . Bengio, P. Haffner, Gradient-based learning applied to document recognition, Proceedings of the IEEE 86 (11) (1998) 2278–2324.doi:10.1109/5.726791

    1998 doi

  39. [47]

    R. A. Fisher, The use of multiple measurements in taxonomic problems, Annals of Eugenics 7 (2) (1936) 179–188.doi:10.1111/j.1469-1809. 1936.tb02137.x

    1936 doi

  40. [48]

    H. T. Nguyen, M. Usman, R. Buyya, Qfaas: A serverless function-as-a-service framework for quantum computing, Future Generation Computer Systems 154 (2024) 281–300.doi:https:// doi.org/10.1016/j.future.2024.01.018. URLhttps://www.sciencedirect.com/ science/article/pii/S0167739X...

    2024 doi

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.