Neural Network Architectures for Scalable Quantum State Tomography: Benchmarking and Memristor-Based Acceleration

Erbing Hua; Heba Abunahla; Jim van Leeuven; King Yiu Yu; Rajendra Bishnoi; Ryoichi Ishihara; Salahuddin Nur; Sebastian Feld; Steven van Ommen

REVIEW 4 major objections 5 minor 39 references

Neural Network Architectures for Scalable Quantum State Tomography: Benchmarking and Memristor-Based Acceleration

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read CNNs and conditional GANs are the most accurate, scalable neural architectures for quantum state tomography, with pure-state infidelity below $2\times 10^{-3}$.

desk verdict A useful but overgeneralized benchmark: the CNN/CGAN ranking holds for noiseless, state-aware inputs, not yet for realistic QST. read the letter →

arxiv 2507.23007 v1 pith:MY2R6OGW submitted 2025-07-30 quant-ph

Erbing Hua , Steven van Ommen , King Yiu Yu , Jim van Leeuven , Rajendra Bishnoi , Heba Abunahla , Salahuddin Nur , Sebastian Feld

show 1 more author

Ryoichi Ishihara

This is my paper · ORCID

classification quant-ph

keywords quantumstatetomographyneuralnetworksconvolutionalconditionalgenerativeadversarialspikingvariationalautoencodermemristorcomputation-in-memorymeasurementbasisscaling

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

Quantum state tomography (QST) reconstructs a quantum state from measurements, but the number of measurement bases grows as $4^N$ with qubit count $N$, so scalability is the central obstacle. The paper benchmarks seven neural architectures — CNN, FCN, RNN, Transformer, CGAN, RBM, and SVAE — under two measurement strategies, expectation values (M1) and full probability distributions (M2), to see which ones keep fidelity high as the system grows. It finds that CNN and CGAN are the most accurate and scalable: on pure states with M1 they reach infidelity below $2\times 10^{-3}$, and among all models they converge fastest and handle noisy mixed states best. The unsupervised spiking variational autoencoder (SVAE) reaches only moderate fidelity, above 0.9 up to about six qubits with $10^5$–$10^6$ shots, but its event-driven design is the most natural fit for low-power neuromorphic hardware. The memristor-based computation-in-memory discussion is explicitly a proposal: no hardware-level simulation or implementation was performed, so the energy and latency gains are estimated rather than measured.

What carries the argument

The load-bearing object is the neural-network measurement-to-state pipeline with a custom DensityMatrix layer: each architecture maps measurement statistics into a reconstructed density matrix, recomputes expectation values from that matrix, and is trained by mean-squared-error (MSE) between reconstructed and true measurements, with CGAN using its adversarial loss instead. The benchmark is driven by two input formats — M1, a single scalar expectation value per measurement basis, and M2, the full probability distribution over outcomes — and by the empirically required number of measurement bases $|M|$ needed to reach fidelity $\approx 0.99$. For the hardware claim, the relevant mechanism is the memristor crossbar, which performs matrix-vector multiplication in analog form by storing network weights as conductances and applying inputs as voltages, so that multiply-accumulate operations happen in place and avoid the traffic bottleneck of conventional separated-memory architectures; the SVAE's spiking, event-driven encoding is what makes it the paper's leading candidate for that kind of acceleration.

What would settle it

Take the same seven architectures and repeat the benchmark with finite-shot inputs — for example, $10^3$, $10^4$, and $10^5$ shots per basis — instead of exact M1 and M2 statistics, and check whether CNN and CGAN still hold lower infidelity than every near competitor; if another architecture wins, or if the SVAE's shot robustness overtakes them, the paper's central ranking fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central result is an architecture ranking for neural-network QST. Convolutional neural networks (CNN) and conditional generative adversarial networks (CGAN) consistently achieve the highest reconstruction fidelity across pure and mixed states, with infidelity below $2\times 10^{-3}$ for pure states under the expectation-value method M1; both also reach fidelity above 0.99 on a noisy 9-qubit mixed state within roughly the same iteration count, while FCN, RNN, and Transformer do not. RBM and RNN are the weak performers: RNN oscillates without converging on a 9-qubit GHZ state and RBM plateaus near fidelity 0.43. The paper also establishes that the measurement strategy matters as much as the architecture: probability-based M2 needs far fewer measurement bases — for an 8-qubit GHZ state, 2 bases instead of 11, and for the W state 12 instead of 130 — but M2 requires many more experimental shots per basis, so M1 is judged more practical for real mixed-state experiments. The SVAE, evaluated separately because it is unsupervised and hardware-oriented, reconstructs pure GHZ states up to about 7 qubits with $10^5$–$10^6$ shots, with fidelity above 0.9 up to about six qubits and declining as qubit number grows.

Load-bearing premise

The ranking assumes that the M1 and M2 training inputs are exact, noiseless expectation values and probabilities computed directly from the true quantum state; if realistic measurement shot noise is added, the reported fidelities and the ordering of architectures could change.

Editorial extensions

If this is right

For high-accuracy reconstruction of pure states, the practical choice is CNN or CGAN with M1, since both stay below $2\times 10^{-3}$ infidelity while converging in far fewer iterations than RNN or RBM.
Probability-based M2 can cut the number of distinct measurement bases dramatically (GHZ needs only 2 bases at 8 qubits), which makes it attractive when each basis can be sampled many times; when shot budget is limited, M1 is the safer strategy.
The SVAE offers a different operating point: moderate fidelity with sparse, event-driven computation, positioning it for embedded or on-chip QST where energy matters more than peak accuracy.
RBM and RNN should be avoided for scalable QST: RNN fails to converge stably and RBM stalls around fidelity 0.43 on the 9-qubit GHZ test.
If memristor crossbars accelerate the matrix-vector multiplications in CNN and CGAN and the spiking operations in SVAE as proposed, the energy cost of reconstruction could drop, but the paper does not yet provide hardware measurements to confirm this.

Reading between the lines

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

Beyond the paper: because the M1 and M2 inputs are exact noiseless statistics, the reported base counts are optimistic; adding realistic shot noise would likely raise $|M|$ and could change the ordering, and the SVAE's own shot-dependent curves show how strong this effect can be.
Beyond the paper: the M2 advantage suggests a promising hybrid that the paper did not test — pair M2's few-base protocol with CNN or CGAN, or combine it with compressed sensing, to push toward few-basis, high-fidelity tomography.
Beyond the paper: if SVAE fidelity can be raised through better encoders or more expressive spiking decoders, its hardware compatibility would make it the preferred architecture for field-deployed quantum diagnostics; the current 6–7 qubit ceiling is the main obstacle.
Beyond the paper: the benchmark's fidelity-per-computation-time framing gives hardware designers a concrete shortlist — CNN and CGAN for accuracy, SVAE for energy — and the missing hardware simulation is the obvious next validation step.

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

Fairly useful paper. The genuinely new piece is the first uniform benchmark of seven NN architectures for QST under two measurement strategies, plus the SVAE-to-QST application. The experiments are described in enough detail to be repeatable in principle, and the authors cite the right prior work. Credit also for the limitations paragraph: they admit the memristor/CiM discussion is not backed by hardware simulation.

The soft spots are real but mostly fixable. The biggest one: for every model except RBM and SVAE, the inputs are exact expectation values and probabilities computed from the true state. No shot noise. So the headline "CNN and CGAN scale more robustly" is only established for ideal measurement statistics. SVAE's own shot sweep shows fidelity collapses below 10^5 shots, which makes the Discussion's claim that SVAE is robust under data-sparse conditions hard to square with Figure 5. Second, the M1 basis-selection step drops zero-expectation bases using knowledge of the true state, so the small |M| values in Table 2 (e.g., 11 bases for 8-qubit GHZ) are not protocol-level numbers; they are post-selected. Third, no error bars, unclear whether "100 iterations" means independent runs, and the code/data links are placeholders ("available at QST data"), so reproducibility cannot be checked.

None of this is a load-bearing mathematical error. The comparison is coherent on its own terms; the problem is that the stated conclusion generalizes from noiseless simulation to practical QST without evidence that the ranking survives finite shots. That is exactly the kind of thing revision can fix: add shot-noise sweeps for all architectures, report variance over seeds, use a basis-selection rule that does not require the true state, and actually link the artifacts.

I would send this to a serious referee. The benchmark is valuable for the QST subfield even if the headline needs qualification, and the issues are addressable. I would not cite the scaling ranking as is, but after the shot-noise revision it would be a useful comparison point. Take to reading group if you want a lively discussion about benchmark protocols.

Referee Report

4 major / 5 minor

Summary. This manuscript benchmarks seven neural network architectures (FCN, CNN, RNN, CGAN, Transformer, RBM, SVAE) for quantum state tomography on pure and mixed states up to 9 qubits, using two idealized measurement strategies: M1 (exact expectation values) and M2 (exact probability distributions). The main claims are that CNN and CGAN achieve the highest fidelities and scale most robustly, while SVAE offers moderate fidelity with attractive hardware compatibility. The paper also discusses memristor-based computation-in-memory (CiM) as a possible acceleration platform, with an explicit caveat that no hardware-level simulations were performed. A separate analysis of SVAE shows fidelity degrading with fewer measurement shots and with more qubits. The manuscript provides code and data availability statements and a detailed appendix of network architectures.

Significance. If the central claim holds, the paper provides a useful practical guideline for choosing neural architectures for QST and identifies SVAE as a candidate for low-power deployment. The benchmark is broad in architectural coverage and includes both pure and mixed states, two measurement formats, and an explicit shot-count study for SVAE. The paper is honest about several limitations, including the absence of hardware-level validation and the degradation of SVAE fidelity beyond six qubits. The code and data availability statements support reproducibility. However, the headline claims rest on idealized noiseless inputs, a state-dependent measurement-basis selection, and an unequal treatment of the RBM baseline, so the practical conclusions are not yet fully supported.

major comments (4)

[Results: Neural Network Performance Evaluation; Methods: M1/M2] The architecture ranking in Figs. 3 and 4 rests on inputs that are exact expectation values (M1) and exact probability distributions (M2) computed from the true state via Eqs. (6)-(9), with no finite-sampling noise for FCN, CNN, CGAN, RNN, and Transformer; only RBM (Appendix B.4) and SVAE (Section 5) receive noisy shot data. Since the paper's own SVAE results (Fig. 5) show fidelity falling sharply as shots decrease, the relative ordering of CNN/CGAN versus the other supervised models is not established under realistic finite-shot conditions. Please repeat the benchmark for all architectures with finite shot counts (e.g., 10^2 to 10^6) and report whether the ranking persists.
[Results: Measurement bases; Table 2] The reported |M| values are obtained by first computing expectation values from the true density matrix and then discarding all bases with zero or near-zero expectation value (Section 'Measurement bases'). This is a state-dependent selection that requires knowledge of the state being reconstructed; a tomography protocol cannot know in advance which Pauli bases will have nonzero expectation value. Consequently, the sub-exponential scaling in |M| claimed in Fig. 1(c) and Table 2 is a property of a curated, state-specific input, not of a generally implementable measurement scheme. Please test with informationally complete or randomly selected bases, or with an adaptive protocol that does not use the true state.
[Appendix B.4 (RBM); Results: Neural Network Performance Evaluation] The RBM is evaluated under different data conditions from the other models: Appendix B.4 states that the RBM 'uses measurement data that inherently includes statistical fluctuations,' while M1/M2 inputs for FCN, CNN, CGAN, RNN, and Transformer are exact. This unequal treatment means the poor RBM performance in Figs. 3 and 4 may be due to shot noise rather than architectural limitations. Please either provide all models with the same noisy measurement statistics or evaluate the RBM in the same noiseless setting, and specify clearly which protocol is used.
[Discussion; Section 5 (SVAE)] The claim that SVAE is 'a strong candidate for embedded, low-power hardware implementations' (Abstract) over-reaches the evidence presented: the Discussion explicitly states that 'hardware-level simulations of memristor-based CiM accelerators have not yet been performed,' and Fig. 5 shows SVAE fidelity not exceeding 0.9 beyond 6 qubits even at 10^6 shots. The hardware-suitability conclusion is therefore speculative. Please soften the claim or provide quantitative energy/latency estimates from actual CiM simulation.

minor comments (5)

[Figure 1 caption] The caption contains a duplicated label '(b) (b) Scaling of Hilbert space size' and should be corrected.
[Results: Neural Network Performance Evaluation] The text states that 'CNN offers the most balanced performance' but earlier says 'CGAN and CNN consistently achieve the minimal infidelity,' with CGAN reportedly the best; please reconcile which architecture is claimed to be most balanced.
[Appendix B] Training hyperparameters (epochs, learning rate, batch size, early-stopping criteria) are reported for SVAE (Tables B8-B9) but not for FCN, CNN, CGAN, RNN, and Transformer; without these details the comparison is hard to reproduce despite code availability.
[Background] The sentence 'The methodology used for reconstructing quantum states... involves training a neural network to map measurement data to a target quantum state' contains grammatical errors and an unresolved display of Eq. (11); please edit for clarity.
[References] References [28] and [33] appear twice (in both arXiv and published forms); these should be unified to avoid duplicate citation entries.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the architecture benchmark and |M| counts are empirical outputs, not inputs in disguise.

full rationale

The paper's central claims are benchmarking outcomes rather than derived predictions. Measurement inputs (M1 expectation values, M2 probabilities) are generated from the target states, each network is trained to map those inputs to a density matrix, and fidelity is computed against the same target states; no reported fidelity, convergence time, or required-basis count is defined in terms of the fitted parameters themselves. The CNN/CGAN ranking is an empirical result of training dynamics and is not forced by the definition of the loss or the data. The M1 basis-selection step (restricting to non-zero expectation bases) is a disclosed, state-dependent preprocessing choice; it may limit how the |M| values transfer to realistic protocols, but it is not an equation that reproduces its own input. The SVAE shot-noise study is explicitly an experimental scaling curve, and the memristor/CiM discussion is explicitly acknowledged as not hardware-validated. No load-bearing self-citation, imported uniqueness theorem, or ansatz-via-citation appears; the cited prior work is used for background and qualitative support only. Accordingly, no specific circular step satisfies the evidentiary requirement of exhibiting a reduction by construction.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central ranking claim rests mainly on the simulation setup (noiseless data, basis selection) rather than on fitted constants. The free parameters are hand-chosen hyperparameters that affect the SVAE and RBM results, but they do not determine the CNN/CGAN ranking.

free parameters (5)

SVAE beta (regularization weight) = 0.819
Hand-chosen hyperparameter for the SVAE loss; central to the SVAE fidelity results.
RBM learning rate = 0.1
Set high initially to avoid being stuck at zero fidelity; a hand-tuned training choice.
RBM hidden units = 2N+1
Chosen by hand; from Appendix B.4.
SVAE hidden size scale = 20n
Architectural scaling chosen by hand; controls SVAE capacity across qubit numbers.
Target fidelity threshold = approximately 0.99
Chosen cutoff for 'high-fidelity' reconstruction in Table 2 and Figure 1(c).

assumptions (4)

domain assumption Measurement data are noiseless exact values for M1 and M2
All models except SVAE are trained on exact expectation values and probabilities computed from the true state, with no sampling noise.
ad hoc to paper Bases with zero expectation value add no useful information and can be excluded
The paper excludes zero-expectation bases for M1 based on empirical instability, which uses knowledge of the true state.
domain assumption Neural network outputs become physical density matrices via the custom DensityMatrix layer
Reconstruction quality depends on this layer enforcing Hermiticity, PSD, and trace 1.
standard math Pauli basis informational completeness requires 4^N measurement settings
Used as the reference scaling for worst-case tomography.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Neural Network Architectures for Scalable Quantum State Tomography: Benchmarking and Memristor-Based Acceleration." pith.science (2026). https://pith.science/paper/MY2R6OGW

@misc{pith2026250723007,
  author       = {Pith},
  title        = {Pith review of: Neural Network Architectures for Scalable Quantum State Tomography: Benchmarking and Memristor-Based Acceleration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MY2R6OGW}},
  note         = {Machine review of arXiv:2507.23007}
}

read the original abstract

Quantum State Tomography (QST) is essential for characterizing and validating quantum systems, but its practical use is severely limited by the exponential growth of the Hilbert space and the number of measurements required for informational completeness. Many prior claims of performance have relied on architectural assumptions rather than systematic validation. We benchmark several neural network architectures to determine which scale effectively with qubit number and which fail to maintain high fidelity as system size increases.To address this, we perform a comprehensive benchmarking of diverse neural architectures across two quantum measurement strategies to evaluate their effectiveness in reconstructing both pure and mixed quantum states. Our results reveal that CNN and CGAN scale more robustly and achieve the highest fidelities, while Spiking Variational Autoencoder (SVAE) demonstrates moderate fidelity performance, making it a strong candidate for embedded, low-power hardware implementations.Recognizing that practical quantum diagnostics will require embedded, energy-efficient computation, we also discuss how memristor-based Computation-in-Memory (CiM) platforms can accelerate these models in hardware, mitigating memory bottlenecks and reducing energy consumption to enable scalable in-situ QST. This work identifies which architectures scale favorably for future quantum systems and lays the groundwork for quantum-classical co-design that is both computationally and physically scalable.

Figures

Figures reproduced from arXiv: 2507.23007 by the authors.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 34 canonical work pages

[1]

Nielsen, M. A. & Chuang, I. L.Quantum computation and quantum information (Cambridge university press, 2010)

2010
[2]

Shor, P. W. Polynomial-time algorithms for prime factorization and discrete logarithms on a quantum computer.SIAM review 41, 303–332 (1999)

work page 1999
[3]

& Gutmann, S

Farhi, E., Goldstone, J. & Gutmann, S. A quantum approximate optimization algorithm. arXiv preprint arXiv:1411.4028(2014)

arXiv 2014
[4]

Universal quantum simulators.Science 273, 1073–1078 (1996)

Lloyd, S. Universal quantum simulators.Science 273, 1073–1078 (1996)

work page 1996
[5]

& Rehacek, J.Quantum state estimationVol

Paris, M. & Rehacek, J.Quantum state estimationVol. 649 (Springer Science & Business Media, 2004)

work page 2004
[6]

T., Becker, S

Gross, D., Liu, Y.-K., Flammia, S. T., Becker, S. & Eisert, J. Quantum state tomography via compressed sensing.Physical review letters105, 150401 (2010)

work page 2010
[7]

Efficient quantum state tomography.Nature communications 1, 149 (2010)

Cramer, M.et al. Efficient quantum state tomography.Nature communications 1, 149 (2010)

work page 2010
[8]

Melnikov, A. A.et al. Active learning machine learns to create new quantum experiments. Proceedings of the National Academy of Sciences115, 1221–1226 (2018). ddd19

work page 2018

Show all 39 references

[9]

Torlai, G.et al.Neural-network quantum state tomography.Nature Physics 14, 447–450 (2018)

2018
[10]

Carrasquilla, J., Torlai, G., Melko, R. G. & Aolita, L. Reconstructing quantum states with generative models.Nature Machine Intelligence1, 155–161 (2019)

2019
[11]

& Gärttner, M

Schmale, T., Reh, M. & Gärttner, M. Efficient quantum state tomography with convolutional neural networks.npj Quantum Information8, 115 (2022)

2022
[12]

T., Brodsky, M., Danaci, O

Lohani, S., Kirby, B. T., Brodsky, M., Danaci, O. & Glasser, R. T. Machine learning assisted quantum state estimation. Machine Learning: Science and Technology 1, 035007 (2020)

2020
[13]

R., Huang, C.-J

Ma, H., Dong, D., Petersen, I. R., Huang, C.-J. & Xiang, G.-Y. Neural net- works for quantum state tomography with constrained measurements.Quantum Information Processing23, 317 (2024)

2024
[14]

J., Carrasquilla, J

Morawetz, S., De Vlugt, I. J., Carrasquilla, J. & Melko, R. G. U (1)-symmetric recurrent neural networks for quantum state reconstruction.Physical Review A 104, 012401 (2021)

2021
[15]

Neville, A.et al.Classical boson sampling algorithms with superior performance to near-term experiments.Nature Physics 13, 1153–1157 (2017)

2017
[16]

& Kockum, A

Ahmed, S., Sánchez Muñoz, C., Nori, F. & Kockum, A. F. Quantum State Tomography with Conditional Generative Adversarial Networks.Physical Review Letters 127, 1–8 (2021)

2021
[17]

& Rabitz, H

Ma, H., Sun, Z., Dong, D., Chen, C. & Rabitz, H. Attention-based trans- former networks for quantum state tomography.arXiv preprint arXiv:2305.05433 (2023)

2023 arXiv
[18]

& Gong, D

Ma, H., Sun, Z., Dong, D. & Gong, D. Learning informative latent representa- tion for quantum state tomography.IEEE Transactions on Emerging Topics in Computational Intelligence(2025)

2025
[19]

& Severini, S

Rocchetto, A., Grant, E., Strelchuk, S., Carleo, G. & Severini, S. Learning hard quantum distributions with variational autoencoders.npj Quantum Information 4, 28 (2018)

2018
[20]

& Situ, H

Chen, C., He, Z., Huang, Z. & Situ, H. Reconstructing a quantum state with a variational autoencoder. International Journal of Quantum Information 19, 2140005 (2021)

2021
[21]

Can programming be liberated from the von neumann style? a func- tional style and its algebra of programs

Backus, J. Can programming be liberated from the von neumann style? a func- tional style and its algebra of programs. Communications of the ACM 21, 613–641 (1978)

1978
[22]

A., Strachan, J

Zidan, M. A., Strachan, J. P. & Lu, W. D. The future of electronics based on memristive systems. Nature electronics1, 22–29 (2018)

2018
[23]

Prezioso, M.et al.Training and operation of an integrated neuromorphic network based on metal-oxide memristors.Nature 521, 61–64 (2015)

2015
[24]

& Wong, H.-S

Ielmini, D. & Wong, H.-S. P. In-memory computing with resistive switching devices. Nature electronics1, 333–343 (2018)

2018
[25]

& Marquardt, F

Krenn, M., Landgraf, J., Foesel, T. & Marquardt, F. Artificial intelligence and machine learning for quantum technologies (2023). ddd20

2023
[26]

D., Williams, R

in Yi, S., Kendall, J. D., Williams, R. S. & Kumar, S. Activity-difference training of deep neural networks using memristor crossbars.Nature Electronics6, 45–51 (2023)

2023
[27]

Kingma, D. P. & Ba, J. Adam: A method for stochastic optimization (2014). URL http://arxiv.org/abs/1412.6980

2014 arXiv
[28]

& Kockum, A

Ahmed, S., Sánchez Muñoz, C., Nori, F. & Kockum, A. F. Quantum state tomog- raphy with conditional generative adversarial networks.Physical review letters 127, 140502 (2021)

2021
[29]

Cai, F. et al. A fully integrated reprogrammable memristor-CMOS system for efficient multiply-accumulate operations.Nature Electronics 2, 290–299 (2019). URL http://dx.doi.org/10.1038/s41928-019-0270-x

2019 doi
[30]

Accurate & Energy-efficient ECG Classification using RRAM based DNN Architecture (2021)

Dash, S. Accurate & Energy-efficient ECG Classification using RRAM based DNN Architecture (2021). URL http://repository.tudelft.nl/

2021
[31]

T., Gross, D., Liu, Y.-K

Flammia, S. T., Gross, D., Liu, Y.-K. & Eisert, J. Quantum tomography via compressed sensing: error bounds, sample complexity and efficient estimators. New Journal of Physics14, 095022 (2012)

2012
[32]

& Deutsch, I

Kalev, A., Kosut, R. & Deutsch, I. Quantum tomography protocols with posi- tivity are compressed sensing protocols. njp quant. inf. 1, 15018 (2015).arXiv preprint arXiv:1502.00536 (2015)

2015 arXiv
[33]

Carrasquilla, J., Torlai, G., Melko, R. G. & Aolita, L. Reconstructing quantum states with generative models (2018). URL http://arxiv.org/abs/1810.10584http: //dx.doi.org/10.1038/s42256-019-0028-1

2018 arXiv
[34]

Yao, P. et al. Fully hardware-implemented memristor convolutional neural network. Nature 577, 641–646 (2020)

2020
[35]

Hardware implementation of memristor-based artificial neural networks

Aguirre, F.et al. Hardware implementation of memristor-based artificial neural networks. Nature communications15, 1974 (2024)

2024
[36]

Spiking neuromorphic chip learns entangled quantum states

Czischek, S.et al. Spiking neuromorphic chip learns entangled quantum states. SciPost Physics 12, 039 (2022)

2022
[37]

Klassert, R., Baumbach, A., Petrovici, M. A. & Gärttner, M. Variational learning of quantum ground states on spiking neuromorphic hardware.Iscience 25 (2022)

2022
[38]

Patel, A., Gaikwad, A., Huang, T., Kockum, A. F. & Abad, T. Selective and efficient quantum state tomography for multi-qubit systems. arXiv preprint arXiv:2503.20979 (2025)

2025
[39]

T., Farooq, A

Ahmad, S. T., Farooq, A. & Shin, H. Self-guided quantum state tomography for limited resources. Scientific Reports 12, 5092 (2022). ddd21

2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Nielsen, M. A. & Chuang, I. L.Quantum computation and quantum information (Cambridge university press, 2010)

2010

[2] [2]

Shor, P. W. Polynomial-time algorithms for prime factorization and discrete logarithms on a quantum computer.SIAM review 41, 303–332 (1999)

work page 1999

[3] [3]

& Gutmann, S

Farhi, E., Goldstone, J. & Gutmann, S. A quantum approximate optimization algorithm. arXiv preprint arXiv:1411.4028(2014)

arXiv 2014

[4] [4]

Universal quantum simulators.Science 273, 1073–1078 (1996)

Lloyd, S. Universal quantum simulators.Science 273, 1073–1078 (1996)

work page 1996

[5] [5]

& Rehacek, J.Quantum state estimationVol

Paris, M. & Rehacek, J.Quantum state estimationVol. 649 (Springer Science & Business Media, 2004)

work page 2004

[6] [6]

T., Becker, S

Gross, D., Liu, Y.-K., Flammia, S. T., Becker, S. & Eisert, J. Quantum state tomography via compressed sensing.Physical review letters105, 150401 (2010)

work page 2010

[7] [7]

Efficient quantum state tomography.Nature communications 1, 149 (2010)

Cramer, M.et al. Efficient quantum state tomography.Nature communications 1, 149 (2010)

work page 2010

[8] [8]

Melnikov, A. A.et al. Active learning machine learns to create new quantum experiments. Proceedings of the National Academy of Sciences115, 1221–1226 (2018). ddd19

work page 2018