Q-MAML: Quantum Model-Agnostic Meta-Learning for Variational Quantum Algorithms

Jeihee Cho; Junyong Lee; Shiho Kim

REVIEW 4 major objections 5 minor 1 cited by

Q-MAML: Quantum Model-Agnostic Meta-Learning for Variational Quantum Algorithms

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Q-MAML claims that a classical Learner network can learn to initialize parameterized quantum circuits so that variational optimization converges quickly and avoids barren-plateau regions.

desk verdict A plausible learned-initialization study that overclaims MAML and barren-plateau mitigation; worth reviewing with serious revisions. read the letter →

arxiv 2501.05906 v1 pith:QGJ6WK7T submitted 2025-01-10 quant-ph cs.LG

Junyong Lee , JeiHee Cho , Shiho Kim This is my paper · ORCID

classification quant-phcs.LG

keywords meta-learningvariationalquantumeigensolverparameterinitializationbarrenplateauHeisenbergXYZHamiltonianmolecularquantum-classicalhybridgradientnorm

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

The paper introduces Q-MAML, a hybrid classical-quantum framework that uses a classical neural network, called the Learner, to predict the initial parameters of a parameterized quantum circuit from a description of the Hamiltonian to be minimized. The Learner is pre-trained on a distribution of Hamiltonians using a meta-objective that sums the variational quantum eigensolver (VQE) cost over tasks, and because the gradient of that combined cost can be pushed back through the circuit measurement, the network weights can be trained like an ordinary neural network. After pre-training, a new Hamiltonian is given to the frozen Learner, and only a few PQC updates are needed to reach an accurate energy. The authors report faster convergence and better final gaps than zero, pi, reduced-domain uniform, and Gaussian initializations on Heisenberg XYZ chains up to 20 qubits and molecular Hamiltonians up to 14 qubits, and they interpret the moderate gradient norms of their initializations as evidence that the method avoids barren-plateau behavior.

What carries the argument

The central machinery is a parameterized classical network $h_W$, called the Learner, which takes a Hamiltonian task vector $\phi_i$ and outputs the initial PQC parameters $\theta = h_W(\phi_i)$. It is trained with the meta-objective $\arg\min_W \sum_{T_i \sim p(T)} l_{T_i}(g(h_W(\phi_i)))$, where $l_{T_i}$ is the VQE energy expectation, $g$ is the parameterized quantum circuit, and the gradient flows back through the entire chain $dl_{T_i}/dW = (\partial l/\partial \langle O\rangle)(\partial \langle O\rangle/\partial \theta)(\partial \theta/\partial W)$. This lets a classical optimizer update the network weights directly from the quantum circuit's cost, replacing MAML's per-task inner gradient steps with a one-shot prediction.

What would settle it

Sample a held-out set of Heisenberg Hamiltonians with couplings drawn uniformly from, say, $[-10,10]$ instead of $[-3,3]$, pretrain the Learner only on the original range, and compare adaptation speed against Gaussian initialization; if the Q-MAML starting points do not converge faster, the claimed generalization across the problem space fails.

Watch

Extended reading notes

Core claim

The authors claim that a classical network can learn the mapping from Hamiltonian descriptors to useful PQC initial parameters, and that this mapping carries across a task space. Concretely, Q-MAML pre-trains the Learner to minimize the sum of VQE costs over sampled Hamiltonians, so its output parameters sit in regions with moderate gradient norms and small energy gaps; during adaptation, the PQC is fine-tuned from those parameters with standard gradient descent. They report faster convergence and lower final gaps than the four standard initialization baselines on Heisenberg chains with 12 and 20 qubits and molecular Hamiltonians with 10 and 14 qubits, and they interpret the moderate gradient norms as evidence that the initialization avoids barren-plateau regions.

Load-bearing premise

The load-bearing premise is that minimizing the average cost at the starting point over the sampled training Hamiltonians is enough to make that starting point adapt quickly under gradient descent on a new Hamiltonian, and that the sampled Hamiltonians represent the problems Q-MAML will be asked to solve.

Editorial extensions

If this is right

A one-time classical pretraining step can absorb most of the exploration, leaving only a short quantum fine-tuning loop for new Hamiltonians in the trained task space.
The same frozen Learner can be reused on unseen Hamiltonians that fall inside the descriptor distribution, without retraining the classical network.
Starting points with moderate gradient norms should make gradient-based VQE updates more stable, reducing wasted quantum-circuit evaluations.
The framework's task-space encoding generalizes to any Hamiltonian family that can be represented by a fixed-length vector, such as coupling coefficients or molecular-Hamiltonian coefficient vectors.

Reading between the lines

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

Because the meta-objective only minimizes the initial cost and never takes an inner gradient step, Q-MAML is closer to an average-cost initialization than to MAML's fast-adaptation objective; testing a version with one inner update would show whether the inner step adds anything.
A strong test of the generalization claim is to evaluate the pretrained Learner on Hamiltonians whose descriptors lie outside the training distribution, for example Heisenberg couplings beyond $J \in [-3,3]$; if performance collapses, the task-space distribution is the real bottleneck.
If the moderate-gradient-norm observation holds across more circuit families, it suggests a landscape-geometry criterion for good initialization: choose points that sit in basins with bounded gradient norms, which could be measured without knowing the ground state.

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.

Desk Editor's Note

read the letter

Quick take: this is a legitimate but over-sold empirical study. The method is a classical network that maps Hamiltonian coefficients to PQC initial parameters, trained by minimizing the average energy over a task distribution. That idea already exists in Meta-VQE (Cervera-Lierta et al., 2021), which the authors cite but never compare against. The 'MAML' label is misleading: Equation 2 is direct loss minimization with no inner gradient step, so it is not model-agnostic meta-learning in the Finn et al. sense. What the paper does well is run the method on two meaningful problem families (Heisenberg XYZ and small molecules), show that the learned initialization converges faster than zero, pi, uniform, and Gaussian starts in 10-20 qubit simulations, and report parameter statistics that indicate the network is doing something task-specific rather than outputting a constant. That is a real, if incremental, result.

The soft spots are proportionate to the claims. The barren-plateau discussion is the weakest part. Figures 6 and 7 show mean gradient norms at the initial point for n=12/20 and n=10/14. Barren plateaus are defined by exponential decay of the variance of gradient components with qubit count, not by the mean norm. The authors even hedge in the text ('Although this observation needs further discussion...'). The stress-test point is exactly right: the pretraining gradient in Eq. 3 contains d/dθ, so if the ansatz were in a barren plateau, the Learner's own training signal would vanish exponentially and the method could not learn. The current system sizes are too small to establish that this failure does not occur, so the conclusion that Q-MAML 'circumvents' barren plateaus is unsupported. The test sets are also small (16 and 6 Hamiltonians), with no statistical significance or final energy error bars, and no code or data is provided. The missing Meta-VQE baseline is the most serious omission: without it, the claimed advantage over 'traditional initialization' does not separate the learned-initialization idea from this particular implementation.

Who is this for? Researchers working on initialization strategies for VQAs might find the empirical evidence worth a look, and the paper is clearly written. It deserves a serious peer review, but the referee should push hard for a rescoped abstract, a Meta-VQE comparison, variance-based BP diagnostics, and release of code and data. As is, I would not cite it in my own work.

Referee Report

4 major / 5 minor

Summary. The paper proposes Q-MAML, a hybrid classical-quantum framework in which a classical neural network (the Learner) maps a Hamiltonian task vector to initial parameters of a parameterized quantum circuit used for variational quantum eigensolver-type optimization. In the pre-training phase the Learner is trained by minimizing a meta-objective that is a sum of VQE cost functions over a distribution of Hamiltonians; in the adaptation phase the PQC is fine-tuned from the Learner-provided initialization. The authors report experiments on Heisenberg XYZ Hamiltonians (12 and 20 qubits) and molecule Hamiltonians (10 and 14 qubits), comparing against Zero, pi, reduced-domain Uniform, and Gaussian initialization methods, and conclude that Q-MAML provides faster convergence and mitigates barren plateaus.

Significance. If the central claims held, the paper would offer a useful heuristic for initializing PQCs across families of Hamiltonians, and the empirical comparison against several standard initialization schemes is a reasonable experimental design. The paper also avoids circularity: the Learner is trained on a task distribution and evaluated on held-out tasks. However, the two headline claims—that the method is MAML and that it circumvents barren plateaus—are not supported by the presented algorithm and evidence. The algorithm in Eq. (2) and Algorithm 1 is not MAML because it contains no inner gradient step, and the barren-plateau discussion rests on mean gradient norms at modest qubit numbers without the variance-scaling analysis that defines barren plateaus. These are load-bearing issues for the title, abstract, and conclusion.

major comments (4)

[Method, Eq. (2) and Algorithm 1] The pretraining objective in Eq. (2), argmin_W sum_i l_Ti(g(h_W(phi_i))), and the update in Algorithm 1 line 10, W <- W + alpha grad_W sum_i l_Ti(...), contain no inner gradient step on the PQC parameters. MAML is defined by optimizing the parameters of the base model so that one or a few gradient updates on a new task lead to low loss; here the Learner output is used directly as the PQC initialization and there is no adaptation loop inside the meta-objective. Consequently, the method is not MAML as claimed in the title and abstract, but rather direct minimization of a combined VQE cost over the task distribution. This is not merely a naming issue: the claimed fast-adaptation mechanism of MAML is not present. The authors should either add a genuine inner-loop adaptation step (with the associated Hessian or implicit-differentiation approximation) and compare against a true MAML baseline, or reframe the method and claims as "learned initialization" rather than MAML.
[Discussion, "A New Perspective on Barren Plateaus", Figures 6 and 7] The barren-plateau claim is not supported by the evidence. Barren plateaus are characterized by the exponential decay of the variance of cost-function gradients with the number of qubits, or by exponential concentration of the cost, not by a moderate mean gradient norm at a single system size. The paper reports only mean gradient norms at n=12/20 (Heisenberg) and n=10/14 (molecule), with no variance statistics, no cost-concentration analysis, and no scaling with n. The text itself concedes, "Although this observation needs further discussion, we can assume..." in the Discussion section. To support the conclusion that Q-MAML "effectively circumvents the common challenges posed by barren plateaus," the authors need a scaling test: compute the variance of gradients (or the variance of the cost) at the Q-MAML initialization as a function of n for a systematically growing family of instances, and show it does not decay exponentially. Otherwise the barren-plateau conclusion should be removed or weakened to a statement about the specific small systems studied.
[Method, Eq. (3)] The pretraining gradient in Eq. (3) contains the factor d<O>/dtheta, i.e., the PQC gradient with respect to its parameters. If the underlying PQC exhibits a barren plateau, this factor vanishes exponentially with the number of qubits, so the Learner's own training signal would vanish in exactly the regime where barren-plateau mitigation is most needed. The experiments at 12 to 20 qubits cannot rule out this failure mode. The authors should either provide variance-based evidence that the PQC gradients in Eq. (3) do not vanish exponentially for the used ansatz, or explicitly restrict the barren-plateau claim to the tested sizes.
[Experiments and Results, Figure 4] The empirical claim of faster and better convergence is based on only 16 held-out Heisenberg Hamiltonians and 6 held-out molecule Hamiltonians, with no multiple seeds, no statistical significance tests, and no error bars on the final gap. The shaded regions in Figure 4 show one standard deviation over tasks, not over independent training runs of the whole Q-MAML procedure. Since the central conclusion depends on Q-MAML outperforming the baselines on average, the authors should report means and standard errors over several random initial-seed repetitions of the Learner pretraining and adaptation, ideally with a paired significance test or confidence intervals.

minor comments (5)

[Algorithm 1, line 10 and Algorithm 2, line 6] Both pseudocode updates use the form theta <- theta + alpha * grad, which is gradient ascent, whereas Eq. (2) and the text state minimization. The updates should be theta <- theta - alpha * grad (or the sign convention should be clarified).
[Figure 1 caption] The notation "P li" in the caption is undefined and appears to be a sum over individual landscape terms; please use an explicit summation symbol or define the operator.
[Abstract and Introduction] There are grammatical errors, e.g., "This approach aim to achieve" in the abstract and "The result implies that the Learner successfully estimates initial parameters that generalize" should be "estimates initial parameters that generalize" is acceptable but "aim" should be "aims".
[Appendix references] The paper repeatedly refers to an Appendix for the Hamiltonian task-space construction, ansatz design, and additional qubit counts, but the provided manuscript does not include the appendix content. Please ensure the appendix is actually included in the submitted version and that all referenced figures and tables appear.
[General] No code or reproducibility details are provided beyond overall experimental settings; adding a data/code availability statement and exact hyperparameters for each experiment would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity identified; the Learner is trained on an initial-cost objective and evaluated on held-out adaptation tasks, with no fitted-input-as-prediction or self-citation reduction.

full rationale

The paper's derivation chain is training a classical Learner to minimize the average initial-point cost (Eq. 2) and then evaluating adaptation performance on held-out Hamiltonians from the same task distribution. The trained quantity (initial cost) and the reported results (adaptation convergence speed and final energy gap in Figures 3 and 4) are distinct, and the held-out evaluation protocol means test results are not used to set constants or refit the Learner. The fact that Eq. 2 omits the inner gradient step of true MAML is a framing and soundness concern rather than a circularity: the training objective is defined independently of the later evaluation metric, so the improvement in adaptation is not forced by construction. The barren-plateau discussion is speculative and under-evidenced because it relies on small-system gradient norms rather than variance scaling, but it is based on measured diagnostics and not on an assumption that already contains the conclusion. The citation to Nemkov, Kiktenko, and Fedorov (2024) supports the differentiability of combined cost functions, and the chain-rule decomposition in Eq. 3 is standard; no self-citation chain or imported uniqueness theorem compels the result. Therefore no circular step reduces the paper's claims to its own inputs.

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

The central claim rests on the task distribution, the ansatz, and chain-rule differentiability; no new physical entities are introduced. The main free choices are the task-space definitions and circuit depths, which are not justified by sensitivity analysis.

free parameters (3)

task distribution range for Heisenberg = Jx, Jy, Jz randomly sampled in [-3, 3] with interval 0.1
The generalization claim depends on this chosen distribution, and no sensitivity analysis is given.
molecule dataset construction = 12 distinct bond lengths, 5 and 7 active orbitals (10 and 14 qubits)
The task space is defined by this specific set; representativeness is assumed.
PQC ansatz depth = up to 20 layers (Heisenberg), 7 layers (Molecule)
The expressibility and trainability of the ansatz depend on these depths; no ablation is shown.

assumptions (4)

domain assumption d<O>/dθ is differentiable and computable via the parameter-shift rule
Invoked in Equation 3 for backpropagation through the quantum circuit.
domain assumption Combining cost functions across tasks is a valid training objective
Stated before Equation 2, citing Nemkov et al. (2024); the paper relies on this to justify the meta-objective.
ad hoc to paper The sampled task distribution is representative of the target problem space
The generalization claim rests on this distributional assumption, with no theoretical or empirical justification beyond the specific experiments.
domain assumption The chosen ansatz can represent the ground states of the tested Hamiltonians
VQE success requires sufficient ansatz expressibility; this is assumed without explicit verification beyond final energies.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Q-MAML: Quantum Model-Agnostic Meta-Learning for Variational Quantum Algorithms." pith.science (2026). https://pith.science/paper/QGJ6WK7T

@misc{pith2026250105906,
  author       = {Pith},
  title        = {Pith review of: Q-MAML: Quantum Model-Agnostic Meta-Learning for Variational Quantum Algorithms},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QGJ6WK7T}},
  note         = {Machine review of arXiv:2501.05906}
}

read the original abstract

In the Noisy Intermediate-Scale Quantum (NISQ) era, using variational quantum algorithms (VQAs) to solve optimization problems has become a key application. However, these algorithms face significant challenges, such as choosing an effective initial set of parameters and the limited quantum processing time that restricts the number of optimization iterations. In this study, we introduce a new framework for optimizing parameterized quantum circuits (PQCs) that employs a classical optimizer, inspired by Model-Agnostic Meta-Learning (MAML) technique. This approach aim to achieve better parameter initialization that ensures fast convergence. Our framework features a classical neural network, called Learner}, which interacts with a PQC using the output of Learner as an initial parameter. During the pre-training phase, Learner is trained with a meta-objective based on the quantum circuit cost function. In the adaptation phase, the framework requires only a few PQC updates to converge to a more accurate value, while the learner remains unchanged. This method is highly adaptable and is effectively extended to various Hamiltonian optimization problems. We validate our approach through experiments, including distribution function mapping and optimization of the Heisenberg XYZ Hamiltonian. The result implies that the Learner successfully estimates initial parameters that generalize across the problem space, enabling fast adaptation.

Figures

Figures reproduced from arXiv: 2501.05906 by the authors.

**Figure 2.** The overall learning path of Q-MAML. The pre-training phase includes an update of the [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Training trajectory of Learner with different Hamiltonian dataset. (a) Heisenberg 12 qubits (b) Heisenberg 20 qubis (c) Molecule 10 qubits (d) Molecule 14 qubits [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Training trajectory of PQC in the adaptation phase. The average gap value is plotted using 16 and 6 different Hamil [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: Statistics of the parameter initialized using Q-MAML, Uniform, and Gaussian. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: Trend of the PQC’s gradient norm during pre-training phase [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

**Figure 8.** Figure 8: Visualization of optimization landscape in Q [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Artificial intelligence for representing and characterizing quantum systems
quant-ph 2025-09 unverdicted novelty 1.0 of 10

A review organizes AI-based quantum system characterization into ML, deep learning, and language model paradigms, covering property prediction and implicit state reconstruction.

Reference graph

Works this paper leans on

25 extracted references · 17 canonical work pages · cited by 1 Pith paper

[1]

, " * write output.state after.block = add.period write newline

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...
[2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...
[3]

Amaro, D.; Rosenkranz, M.; Fitzpatrick, N.; Hirano, K.; and Fiorentini, M. 2022. A case study of variational quantum algorithms for a job shop scheduling problem. EPJ Quantum Technology, 9(1): 5

work page 2022
[4]

L.; Cerezo, M.; Sone, A.; and Coles, P

Beckey, J. L.; Cerezo, M.; Sone, A.; and Coles, P. J. 2022. Variational quantum algorithm for estimating the quantum fisher information. Physical Review Research, 4(1): 013083

work page 2022
[5]

C.; Endo, S.; Fujii, K.; McClean, J

Cerezo, M.; Arrasmith, A.; Babbush, R.; Benjamin, S. C.; Endo, S.; Fujii, K.; McClean, J. R.; Mitarai, K.; Yuan, X.; Cincio, L.; et al. 2021. Variational quantum algorithms. Nature Reviews Physics, 3(9): 625--644

2021
[6]

S.; and Aspuru-Guzik, A

Cervera-Lierta, A.; Kottmann, J. S.; and Aspuru-Guzik, A. 2021. Meta-Variational Quantum Eigensolver: Learning Energy Profiles of Parameterized Hamiltonians for Quantum Simulation. PRX Quantum, 2: 020329

work page 2021
[7]

Finn, C.; Abbeel, P.; and Levine, S. 2017. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, 1126--1135. PMLR

2017
[8]

Friedrich, L.; and Maziero, J. 2022. Avoiding barren plateaus with classical deep neural networks. Physical Review A, 106(4): 042433

work page 2022

Show all 25 references

[9]

Glorot, X.; and Bengio, Y. 2010. Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, 249--256. JMLR Workshop and Conference Proceedings

2010
[10]

D.; Temme, K.; Harrow, A

Havl \' c ek, V.; C \'o rcoles, A. D.; Temme, K.; Harrow, A. W.; Kandala, A.; Chow, J. M.; and Gambetta, J. M. 2019. Supervised learning with quantum-enhanced feature spaces. Nature, 567(7747): 209--212

2019
[11]

Hochreiter, S.; Bengio, Y.; Frasconi, P.; Schmidhuber, J.; et al. 2001. Gradient flow in recurrent nets: the difficulty of learning long-term dependencies

2001
[12]

Holmes, Z.; Sharma, K.; Cerezo, M.; and Coles, P. J. 2022. Connecting ansatz expressibility to gradient magnitudes and barren plateaus. PRX Quantum, 3(1): 010313

2022
[13]

Huang, H.-L.; Du, Y.; Gong, M.; Zhao, Y.; Wu, Y.; Wang, C.; Li, S.; Liang, F.; Lin, J.; Xu, Y.; et al. 2021. Experimental quantum generative adversarial networks for image generation. Physical Review Applied, 16(2): 024051

2021
[14]

J.; and Rung, T

Jaksch, D.; Givi, P.; Daley, A. J.; and Rung, T. 2023. Variational quantum algorithms for computational fluid dynamics. AIAA journal, 61(5): 1885--1894

2023
[15]

Lubasch, M.; Joo, J.; Moinier, P.; Kiffner, M.; and Jaksch, D. 2020. Variational quantum algorithms for nonlinear problems. Physical Review A, 101(1): 010301

2020
[16]

R.; Boixo, S.; Smelyanskiy, V

McClean, J. R.; Boixo, S.; Smelyanskiy, V. N.; Babbush, R.; and Neven, H. 2018. Barren plateaus in quantum neural network training landscapes. Nature communications, 9(1): 4812

2018
[17]

A.; Kiktenko, E

Nemkov, N. A.; Kiktenko, E. O.; and Fedorov, A. K. 2024. Barren plateaus are swamped with traps. arXiv preprint arXiv:2405.05332

2024 arXiv
[18]

Pascanu, R.; Mikolov, T.; and Bengio, Y. 2013. On the difficulty of training recurrent neural networks. In International conference on machine learning, 1310--1318. Pmlr

2013
[19]

J.; Aspuru-Guzik, A.; and O’brien, J

Peruzzo, A.; McClean, J.; Shadbolt, P.; Yung, M.-H.; Zhou, X.-Q.; Love, P. J.; Aspuru-Guzik, A.; and O’brien, J. L. 2014. A variational eigenvalue solver on a photonic quantum processor. Nature communications, 5(1): 4213

2014
[20]

N.; Sauvage, F.; Kemper, A

Ragone, M.; Bakalov, B. N.; Sauvage, F.; Kemper, A. F.; Marrero, C. O.; Larocca, M.; and Cerezo, M. 2023. A unified theory of barren plateaus for deep parametrized quantum circuits. arXiv preprint arXiv:2309.09342

2023 arXiv
[21]

M.; and Wiebe, N

Schuld, M.; Bocharov, A.; Svore, K. M.; and Wiebe, N. 2020. Circuit-centric quantum classifiers. Physical Review A, 101(3): 032308

2020
[22]

Wang, H.; Wang, Y.; Sun, R.; and Li, B. 2022. Global convergence of maml and theory-inspired neural architecture search for few-shot learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 9797--9808

2022
[23]

Wang, S.; Fontana, E.; Cerezo, M.; Sharma, K.; Sone, A.; Cincio, L.; and Coles, P. J. 2021. Noise-induced barren plateaus in variational quantum algorithms. Nature communications, 12(1): 6961

2021
[24]

Wang, Y.; Qi, B.; Ferrie, C.; and Dong, D. 2023. Trainability enhancement of parameterized quantum circuits via reduced-domain parameter initialization. arXiv preprint arXiv:2302.06858

2023 arXiv
[25]

Zhang, K.; Liu, L.; Hsieh, M.-H.; and Tao, D. 2022. Escaping from the barren plateau via gaussian initializations in deep variational quantum circuits. Advances in Neural Information Processing Systems, 35: 18612--18627

2022

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

, " * write output.state after.block = add.period write newline

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

[2] [2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

[3] [3]

Amaro, D.; Rosenkranz, M.; Fitzpatrick, N.; Hirano, K.; and Fiorentini, M. 2022. A case study of variational quantum algorithms for a job shop scheduling problem. EPJ Quantum Technology, 9(1): 5

work page 2022

[4] [4]

L.; Cerezo, M.; Sone, A.; and Coles, P

Beckey, J. L.; Cerezo, M.; Sone, A.; and Coles, P. J. 2022. Variational quantum algorithm for estimating the quantum fisher information. Physical Review Research, 4(1): 013083

work page 2022

[5] [5]

C.; Endo, S.; Fujii, K.; McClean, J

Cerezo, M.; Arrasmith, A.; Babbush, R.; Benjamin, S. C.; Endo, S.; Fujii, K.; McClean, J. R.; Mitarai, K.; Yuan, X.; Cincio, L.; et al. 2021. Variational quantum algorithms. Nature Reviews Physics, 3(9): 625--644

2021

[6] [6]

S.; and Aspuru-Guzik, A

Cervera-Lierta, A.; Kottmann, J. S.; and Aspuru-Guzik, A. 2021. Meta-Variational Quantum Eigensolver: Learning Energy Profiles of Parameterized Hamiltonians for Quantum Simulation. PRX Quantum, 2: 020329

work page 2021

[7] [7]

Finn, C.; Abbeel, P.; and Levine, S. 2017. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, 1126--1135. PMLR

2017

[8] [8]

Friedrich, L.; and Maziero, J. 2022. Avoiding barren plateaus with classical deep neural networks. Physical Review A, 106(4): 042433

work page 2022