Pith. sign in

REVIEW 4 major objections 6 minor 35 references

Perturbative Gradient Training: A novel training paradigm for bridging the gap between deep neural networks and physical reservoir computing

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Perturbative Gradient Training approximates gradients from two forward passes, letting black-box physical reservoirs be trained at any depth in a neural network.

desk verdict Honest proof-of-concept that gets oversold: the update rule is SPSA in disguise, and the hardware evidence never shows training through the reservoir to pre-reservoir layers. read the letter →

arxiv 2506.04523 v1 pith:LLHORFJF submitted 2025-06-05 cs.LG cond-mat.mes-hallcs.ETcs.NEphysics.comp-ph

Cliff B. Abbott , Mark Elo , Dmytro A. Bozhko This is my paper
classification cs.LGcond-mat.mes-hallcs.ETcs.NEphysics.comp-ph
keywords perturbativegradienttrainingreservoircomputingphysicalmagnonicauto-oscillationringbackpropagation-freeforward-onlyestimationenergy-efficientAIspin-wave
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 Perturbative Gradient Training (PGT), a way to train neural networks that contain physical reservoir computers, whose internal dynamics are a black box and cannot be traversed by backpropagation. PGT estimates a gradient from just two forward passes: one with a small random perturbation added to the parameters, one with it subtracted, and uses that scalar estimate to update every parameter via a standard optimizer. The authors demonstrate the method on a dense network, on a transformer whose feed-forward layers are replaced by reservoirs, and on experimental hardware built around a magnonic auto-oscillation ring. Their central claim is that PGT reaches performance comparable to backpropagation in exactly the situations where backpropagation is unavailable, which matters because physical reservoirs could cut AI energy costs substantially if they could be embedded at arbitrary depth.

What carries the argument

The load-bearing object is the random perturbation matrix $[PM]$: a matrix of integers in the range $[-r,r]$, optionally thinned by a dropout probability, which defines a single random direction in parameter space. The update is built from the scalar finite-difference ratio $G$ via $G \cdot [PM] / [\mathrm{Counts}]$, where $[\mathrm{Counts}]$ tracks each parameter's participation so the step size is scaled correctly. Because only two forward passes are needed per sample, the cost of the gradient estimate is independent of the number of parameters in the reservoir, which is what makes training through a black-box physical device feasible.

What would settle it

On a small network with a known backpropagation gradient, compute the PGT update matrix for many random perturbation matrices and typical settings of the range $r$, dropout, and scale $\delta$. If the mean cosine similarity between the PGT update and the true gradient is not clearly positive, or if smaller $\delta$ does not reduce the variance of the estimate, then the reported convergence cannot be attributed to an accurate gradient approximation.

Watch

Extended reading notes

Core claim

Perturbative Gradient Training (PGT) replaces backpropagation with two forward passes. For each training sample the algorithm adds a random integer perturbation matrix $[PM]$ to the parameters to get $\theta_{p+}$, subtracts it to get $\theta_{p-}$, computes the loss for both, and forms the scalar gradient estimate $G = (L(\theta_{p+}) - L(\theta_{p-}))/(2\delta)$. It then forms the update matrix $G \cdot [PM] / [\mathrm{Counts}]$ and hands it to a standard optimizer such as SGD or Adam. The paper claims, and demonstrates on a dense network, a transformer with a reservoir layer, and a physical magnonic auto-oscillation ring, that this estimate is good enough to train networks in which the reservoir is a black box, achieving performance comparable to backpropagation where backpropagation cannot be applied.

Load-bearing premise

The method bets that one random nudge up and one random nudge down in parameter space, combined with how much each weight participated in that nudge, tells a standard optimizer enough about which direction to move every weight to keep converging.

Editorial extensions

If this is right

  • Physical reservoirs can be placed at arbitrary depth in a neural network, not just at the input, because no gradient needs to flow through the reservoir.
  • Training cost per sample is exactly two forward passes, independent of the number of reservoir parameters, so deeper reservoirs do not make training proportionally more expensive.
  • On the dense network, PGT with SGD or Adam reached the SGD backpropagation floor faster than SGD backprop itself, and occasional runs broke through toward the Adam floor.
  • On the reservoir transformer, PGT reached a minimum loss of 1.097 after 745 epochs, about 2.86 times more epochs than Adam backpropagation, and the physical reservoir converged at a similar or slightly faster rate than the simulated one.
  • If the paper's energy estimates hold, replacing electronic layers with physical reservoirs would need only a 7.5 to 35 percent improvement in overall efficiency to break even, making this a concrete route to reducing AI training and deployment energy.

Reading between the lines

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

  • The paper leaves the gradient estimate's quality unmeasured; comparing the PGT update direction with the true backpropagation gradient on a small network would directly show how much of the signal is gradient information rather than noise.
  • Because the update uses one random direction per sample, PGT behaves like a stochastic directional derivative; averaging several perturbations per update would cut variance at the cost of more forward passes, a trade-off the paper does not explore.
  • The same two-forward-pass recipe should transfer to any non-differentiable hardware layer, such as optical or analog electronic devices, so the idea is not tied to magnonic reservoirs.
  • The paper reports that in the dense network nothing changed when only the pre-reservoir layers were trained, so the transformer experiment is the main evidence that gradients reach early layers; a task that requires deep credit assignment would test that directly.
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 / 6 minor

Summary. The paper introduces Perturbative Gradient Training (PGT), a forward-only training method that estimates a gradient along a single random direction in parameter space using two forward passes and a perturbation matrix. The authors test PGT on a simulated dense network and a simulated transformer with a reservoir layer, and on a physical magnonic auto-oscillation ring used as a reservoir. They claim that PGT eliminates the need for backpropagation and enables physical reservoirs to be integrated into neural networks regardless of depth or location, with performance comparable to standard backpropagation in settings where backpropagation is impractical or impossible.

Significance. If substantiated, PGT would be a meaningful step toward training networks containing physical reservoirs without backpropagation, potentially improving energy efficiency in AI training. The paper's strengths include the use of real hardware (a magnonic auto-oscillation ring) and a candid discussion of limitations, including the explicit admission that the small physical network's training occurred only in the readout layer. However, the evidence presented is preliminary and partially contradicts the headline claim: the simulated transformer results show PGT is slower and reaches a worse loss than Adam, and the physical experiments lack the ablations needed to demonstrate that PGT actually trains parameters before the reservoir. The absence of error bars, repeated trials, and a theoretical analysis of the gradient estimator further weakens the support for the central claim.

major comments (4)
  1. [Section IV.C] The paper explicitly states that for the small physical dense network, "there was no change in the performance when training only the portion of the network before the reservoir," implying that all learning occurred in the readout layer. This is exactly the regime PGT is intended to overcome. The physical transformer experiment (Fig. 8) lacks the analogous check: no comparison with pre-reservoir parameters frozen, no per-layer update statistics, and no repeated hardware runs. Consequently, the Abstract and Section I claim that PGT enables physical reservoirs to be integrated "regardless of depth and location of the RC" is unsupported by the experimental evidence, which is a load-bearing issue for the paper's main contribution.
  2. [Section III, Fig. 5] The reported results undermine the "comparable performance" claim. On the transformer, PGT achieved a minimum loss of 1.097 after 745 epochs, taking 2.86x longer, while Adam achieved 0.86 after about 260 epochs. On the dense network, PGT generally plateaued at the SGD limit of 0.2338 while Adam reached 0.0221. No error bars, seeds, or repeated runs are reported, so it is unclear whether the occasional PGT+Adam breakthrough is reproducible. Since the abstract claims comparability with standard backpropagation, these results need to be either supported by repeated trials or reframed in the main claims.
  3. [Section II, Eqs. (4)-(9)] The gradient estimator in Eq. (8) is a single finite-difference ratio along one random direction, weighted by the perturbation matrix in Eq. (9). The paper provides no analysis of the bias or variance of this estimator, no conditions on delta, r, or dropout for the approximation to be faithful, and no convergence guarantees for the resulting SGD/Adam updates. Given that dropout rates of 0.999 and 0.9999 are needed for the transformer, the relationship between perturbation sparsity and estimator quality should be analyzed or ablated; without this, the method's behavior is not adequately explained.
  4. [Section III and IV] Hyperparameters, especially dropout scale, perturbation range r, delta, and learning rate, appear to have been selected on the same datasets used to report success, without a held-out tuning procedure or sensitivity analysis. For example, Section III reports that only dropout 0.9999 "consistently train[ed]" the transformer. This makes it difficult to assess generalization and risks overfitting to the test set; the authors should provide a clear tuning protocol and report results across multiple hyperparameter settings.
minor comments (6)
  1. [Section II] In the sentence "we analyze how that change effected the loss," "effected" should be "affected."
  2. [Section IV.A] The phrase "gets it's name" should be "gets its name."
  3. [Section II, Eq. (9)] The [Counts] matrix is not precisely defined; please specify how it is computed from the perturbation matrix and its exact role in scaling the update.
  4. [Figure 1] The axes and parameter-space directions in Figure 1 are not labeled, making the claim of "8 directions" for r=1 and "16 directions" for r=2 difficult to verify.
  5. [Section IV.B] The reservoir characterization reports C_STM=2.91 and C_PC=0.01, but there is no discussion of how these values relate to the requirements of the transformer task; adding such a discussion would help readers judge the experimental setup.
  6. [General] The paper does not provide a reproducibility statement: no code, no detailed hyperparameter settings (beyond a few values), and no description of data preprocessing (e.g., tokenization and embedding initialization for Multi30k) are given.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: PGT is an explicit forward-pass estimator, and the main weaknesses are missing ablations and empirical support, not circular reasoning.

full rationale

The paper's core derivation chain is the definition of PGT itself: Eqs. (4)-(9) specify a finite-difference estimate along a randomly perturbed direction, Grad = (Lp+ - Lp-)/(2*delta), and form an update by multiplying by the perturbation matrix and scaling by counts. This is an estimator defined directly from forward-pass losses; it is not derived from a target result or fitted to the data it later claims to predict. There is no self-definitional step, no fitted parameter renamed as a prediction, and no imported uniqueness theorem. The one self-citation, Ref. [19] by two of the current authors, appears in a broad citation list [10]-[20] supporting energy-savings estimates; it is not load-bearing for the method's derivation or for the central training claim. The paper's own admission in Section IV.C that the physical dense-network experiment trained only the readout layer ('there was no change in the performance when training only the portion of the network before the reservoir... all the training was occurring in the readout layer post reservoir') is a limitation on the strength of the empirical claim, not a circularity: it does not make the PGT update equivalent to its inputs. Similarly, the high dropout rates used for the transformer were tuned on the same task, which is a selection or overfitting risk, but the paper does not present a held-out quantity as a prediction derived from those tuned values. The comparison with backpropagation is external and falsifiable, even if incomplete. Thus the derivation is self-contained and the circularity score is 0.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

PGT introduces no new physical entities. It rests on unproved smoothness and estimator assumptions, and on hyperparameters tuned per task. The physical reservoir's repeatability and nonlinearity are asserted but not quantitatively characterized beyond C_STM and C_PC.

free parameters (5)
  • dropout scale = 0.999 (dense net), 0.9999 (transformer)
    Chosen per task to make PGT train; explicitly described as necessary for breakthrough on dense net and for consistent transformer training.
  • perturbation range r = not reported
    Hyperparameter controlling the set of possible perturbation directions; no ablation is reported.
  • perturbation scaling factor delta = not reported
    Step size for the finite difference in Eqs. 4-8; no value or sensitivity study is given.
  • learning rate = not reported
    Standard optimizer hyperparameter; values are not given even though convergence depends on them.
  • reservoir operating point = operated just below auto-oscillation threshold
    Sets the memory depth of the physical reservoir; C_STM=2.91 is reported but no control analysis is provided.
assumptions (3)
  • standard math The finite-difference quotient (L(theta+delta d)-L(theta-delta d))/(2 delta) approximates the directional derivative of the loss for small delta.
    Invoked in Eq. 8; requires a smooth loss and small delta, but no bounds or error analysis are given.
  • domain assumption The random perturbation matrix entries are independent and zero-mean so that the perturbation-weighted update is a usable gradient estimator.
    Invoked implicitly in Eq. 9; no proof of unbiasedness or variance behavior is provided.
  • domain assumption The magnonic auto-oscillation ring is a deterministic, repeatable nonlinear map whose responses can be used as a differentiable surrogate layer.
    The physical reservoir is treated as a black box; no noise characterization or calibration repeatability is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Perturbative Gradient Training: A novel training paradigm for bridging the gap between deep neural networks and physical reservoir computing." pith.science (2026). https://pith.science/paper/LLHORFJF

@misc{pith2026250604523,
  author       = {Pith},
  title        = {Pith review of: Perturbative Gradient Training: A novel training paradigm for bridging the gap between deep neural networks and physical reservoir computing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LLHORFJF}},
  note         = {Machine review of arXiv:2506.04523}
}
read the original abstract

We introduce Perturbative Gradient Training (PGT), a novel training paradigm that overcomes a critical limitation of physical reservoir computing: the inability to perform backpropagation due to the black-box nature of physical reservoirs. Drawing inspiration from perturbation theory in physics, PGT uses random perturbations in the network's parameter space to approximate gradient updates using only forward passes. We demonstrate the feasibility of this approach on both simulated neural network architectures, including a dense network and a transformer model with a reservoir layer, and on experimental hardware using a magnonic auto-oscillation ring as the physical reservoir. Our results show that PGT can achieve performance comparable to that of standard backpropagation methods in cases where backpropagation is impractical or impossible. PGT represents a promising step toward integrating physical reservoirs into deeper neural network architectures and achieving significant energy efficiency gains in AI training.

Figures

Figures reproduced from arXiv: 2506.04523 by the authors.

Figure 1
Figure 1. In a 2-D space, r = 1 corresponds to 8 directions and r = 2 corresponds to 16 possible directions, demonstrating how changing the range effects the possible directions that PGT can search during training. Next, we add (subtract) the perturbations to the parameters and determine the new loss. θp+ = θ + [PM] · δ (4) θp− = θ − [PM] · δ (5) Where [PM] is the perturbation matrix and δ is the scaling factor. A new loss is… view at source ↗
Figure 2
Figure 2. Initial simple network design. Actual sizes: Input 30, 1st Hidden 200, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Transformer Architecture with Feed Forward networks replaced by [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: Performance of PGT on a simulated reservoir transformer. While [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 7
Figure 7. Figure 7: Results for training the initial neural network design with a physical [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]
Figure 8
Figure 8. Figure 8: Performance of the physical implementation of the reservoir trans [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 20 canonical work pages

  1. [1]

    Sustainable ai: Environmental implications, challenges and opportunities,

    C.-J. Wu, R. Raghavendra, U. Gupta, B. Acun, N. Ardalani, K. Maeng, G. Chang, F. Aga, J. Huang, C. Baiet al., “Sustainable ai: Environmental implications, challenges and opportunities,”Proceedings of Machine Learning and Systems, vol. 4, pp. 795–813, 2022

    work page 2022

  2. [2]

    OpenAI’s CEO Says the Age of Giant AI Models Is Already Over — wired.com,

    W. Knight, “OpenAI’s CEO Says the Age of Giant AI Models Is Already Over — wired.com,” https://www.wired.com/story/openai-ceo- sam-altman-the-age-of-giant-ai-models-is-already-over, [Accessed 14- 03-2025]

    work page 2025

  3. [3]

    Hybrid heterogeneous clusters can lower the energy consumption of llm inference workloads,

    G. Wilkins, S. Keshav, and R. Mortier, “Hybrid heterogeneous clusters can lower the energy consumption of llm inference workloads,” inThe 15th ACM International Conference on Future and Sustainable Energy Systems, ser. e-Energy ’24. ACM, May 2024, p. 506–513. [Online]. Available: http://dx.doi.org/10.1145/3632775.3662830

    work page arXiv 2024

  4. [4]

    From words to watts: Benchmarking the energy costs of large language model inference,

    S. Samsi, D. Zhao, J. McDonald, B. Li, A. Michaleas, M. Jones, W. Bergeron, J. Kepner, D. Tiwari, and V . Gadepally, “From words to watts: Benchmarking the energy costs of large language model inference,” in2023 IEEE High Performance Extreme Computing Conference (HPEC). IEEE, Sep. 2023, p. 1–9. [Online]. Available: http://dx.doi.org/10.1109/HPEC58863.2023...

    arXiv 2023

  5. [5]

    The environmental impact of ai: A case study of water consumption by chat gpt,

    A. George, A.S.Hovan George, and A.S.Gabrio Martin, “The environmental impact of ai: A case study of water consumption by chat gpt,” 2023. [Online]. Available: https://zenodo.org/record/7855594

    work page arXiv 2023

  6. [6]

    Q&A: UW researcher discusses just how much energy ChatGPT uses — washington.edu,

    “Q&A: UW researcher discusses just how much energy ChatGPT uses — washington.edu,” https://www.washington.edu/news/2023/07/ 27/how-much-energy-does-chatgpt-use, [Accessed 14-03-2025]

    work page 2023

  7. [7]

    Reservoir computing approaches to recurrent neural network training,

    M. Luko ˇseviˇcius and H. Jaeger, “Reservoir computing approaches to recurrent neural network training,”Computer Science Review, vol. 3, no. 3, p. 127–149, Aug. 2009. [Online]. Available: http://dx.doi.org/10.1016/j.cosrev.2009.03.005

    doi:10.1016/j.cosrev.2009.03.005 2009

  8. [8]

    Reservoir transformers,

    S. Shen, A. Baevski, A. S. Morcos, K. Keutzer, M. Auli, and D. Kiela, “Reservoir transformers,” 2020. [Online]. Available: https://arxiv.org/abs/2012.15045

    arXiv 2020

Show all 35 references
  1. [9]

    Analogue and physical reservoir computing using water waves: Applications in power engineering and beyond,

    I. S. Maksymov, “Analogue and physical reservoir computing using water waves: Applications in power engineering and beyond,” Energies, vol. 16, no. 14, p. 5366, Jul. 2023. [Online]. Available: http://dx.doi.org/10.3390/en16145366

    2023 doi

  2. [10]

    Advances in coherent magnonics,

    P. Pirro, V . I. Vasyuchka, A. A. Serga, and B. Hillebrands, “Advances in coherent magnonics,”Nature Reviews Materials, vol. 6, no. 12, p. 1114–1135, Jul. 2021. [Online]. Available: http://dx.doi.org/10.1038/s41578-021-00332-w

    2021 doi

  3. [11]

    Hybrid quantum systems based on magnonics,

    D. Lachance-Quirion, Y . Tabuchi, A. Gloppe, K. Usami, and Y . Nakamura, “Hybrid quantum systems based on magnonics,”Applied Physics Express, vol. 12, no. 7, p. 070101, Jun. 2019. [Online]. Available: http://dx.doi.org/10.7567/1882-0786/AB248D

    2019 doi

  4. [12]

    Towards magnonic devices based on voltage- controlled magnetic anisotropy,

    B. Rana and Y . Otani, “Towards magnonic devices based on voltage- controlled magnetic anisotropy,”Communications Physics, vol. 2, no. 1, Aug. 2019. [Online]. Available: http://dx.doi.org/10.1038/s42005-019- 0189-6 7

    2019 doi

  5. [13]

    Magnonic crystals for data processing,

    A. V . Chumak, A. A. Serga, and B. Hillebrands, “Magnonic crystals for data processing,”Journal of Physics D: Applied Physics, vol. 50, no. 24, p. 244001, May 2017. [Online]. Available: http://dx.doi.org/10.1088/1361-6463/aa6a65

    2017 doi

  6. [14]

    A. V . Chumak,Magnon Spintronics. CRC Press, May 2019, p. 247–302. [Online]. Available: http://dx.doi.org/10.1201/9780429423079-6

    2019 doi

  7. [15]

    Nonlinear spin wave coupling in adjacent magnonic crystals,

    A. V . Sadovnikov, E. N. Beginin, M. A. Morozova, Y . P. Sharaevskii, S. V . Grishin, S. E. Sheshukova, and S. A. Nikitov, “Nonlinear spin wave coupling in adjacent magnonic crystals,”Applied Physics Letters, vol. 109, no. 4, Jul. 2016. [Online]. Available: http://dx.doi.org/1...

    2016 doi

  8. [16]

    Nanoscale spin-wave circuits based on engineered reconfigurable spin-textures,

    E. Albisetti, D. Petti, G. Sala, R. Silvani, S. Tacchi, S. Finizio, S. Wintz, A. Cal `o, X. Zheng, J. Raabe, E. Riedo, and R. Bertacco, “Nanoscale spin-wave circuits based on engineered reconfigurable spin-textures,” Communications Physics, vol. 1, no. 1, Sep. 2018. [Online]. ...

    2018 doi

  9. [17]

    Spin wave normalization toward all magnonic circuits,

    A. N. Mahmoud, F. Vanderveken, C. Adelmann, F. Ciubotaru, S. Cotofana, and S. Hamdioui, “Spin wave normalization toward all magnonic circuits,”IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 68, no. 1, p. 536–549, Jan. 2021. [Online]. Available: http://dx.doi...

    2021

  10. [18]

    A switchable spin-wave signal splitter for magnonic networks,

    F. Heussner, A. A. Serga, T. Br ¨acher, B. Hillebrands, and P. Pirro, “A switchable spin-wave signal splitter for magnonic networks,”Applied Physics Letters, vol. 111, no. 12, Sep. 2017. [Online]. Available: http://dx.doi.org/10.1063/1.4987007

    2017 doi

  11. [19]

    Hybrid magnonic reservoir computing,

    C. B. Abbott and D. A. Bozhko, “Hybrid magnonic reservoir computing,” 2024. [Online]. Available: https://arxiv.org/abs/2405.09542

    2024 arXiv

  12. [20]

    Iono–magnonic reservoir computing with chaotic spin wave interference manipulated by ion-gating,

    W. Namiki, D. Nishioka, Y . Nomura, T. Tsuchiya, K. Yamamoto, and K. Terabe, “Iono–magnonic reservoir computing with chaotic spin wave interference manipulated by ion-gating,”Advanced Science, vol. 12, no. 3, Nov. 2024. [Online]. Available: http://dx.doi.org/ 10.1002/advs.202411777

    2024 doi

  13. [21]

    On the importance of initialization and momentum in deep learning,

    I. Sutskever, J. Martens, G. Dahl, and G. Hinton, “On the importance of initialization and momentum in deep learning,” inProceedings of the 30th International Conference on Machine Learning (ICML-13), 2013, pp. 1139–1147

    2013

  14. [22]

    Understand- ing deep learning requires rethinking generalization,

    C. Zhang, S. Bengio, M. Hardt, B. Recht, and O. Vinyals, “Understand- ing deep learning requires rethinking generalization,” inInternational Conference on Learning Representations (ICLR), 2017

    2017

  15. [23]

    Variational quantum algorithms,

    M. Cerezo, A. Arrasmith, R. Babbush, S. C. Benjamin, S. Endo, K. Fujii, J. R. McClean, K. Mitarai, X. Yuan, L. Cincio, and P. J. Coles, “Variational quantum algorithms,”Nature Reviews Physics, vol. 3, no. 9, p. 625–644, Aug. 2021. [Online]. Available: http://dx.doi.org/10.1038...

    2021 doi

  16. [24]

    Breast cancer wisconsin (diagnostic),

    O. M. William Wolberg, “Breast cancer wisconsin (diagnostic),” 1993. [Online]. Available: https://archive.ics.uci.edu/dataset/17

    1993

  17. [25]

    Bottou,Large-Scale Machine Learning with Stochastic Gradient Descent

    L. Bottou,Large-Scale Machine Learning with Stochastic Gradient Descent. Physica-Verlag HD, 2010, p. 177–186. [Online]. Available: http://dx.doi.org/10.1007/978-3-7908-2604-3 16

    2010 doi

  18. [26]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,”

  19. [27]

    Multi30k: Multilingual english-german image descriptions,

    D. Elliott, S. Frank, K. Sima’an, and L. Specia, “Multi30k: Multilingual english-german image descriptions,” inProceedings of the 5th Workshop on Vision and Language. Association for Computational Linguistics,

  20. [28]

    Numerical simulations of a magnonic reservoir computer,

    S. Watt and M. Kostylev, “Numerical simulations of a magnonic reservoir computer,”Journal of Applied Physics, vol. 135, no. 2, Jan

  21. [29]

    Implementing a magnonic reservoir computer model based on time-delay multiplex- ing,

    S. Watt, M. Kostylev, A. B. Ustinov, and B. A. Kalinikos, “Implementing a magnonic reservoir computer model based on time-delay multiplex- ing,”Physical Review Applied, vol. 15, no. 6, Jun. 2021. [Online]. Available: http://dx.doi.org/10.1103/PhysRevApplied.15.064060

    2021 doi

  22. [30]

    A current- controlled magnonic reservoir for physical reservoir computing,

    A. B. Ustinov, R. V . Haponchyk, and M. Kostylev, “A current- controlled magnonic reservoir for physical reservoir computing,” Applied Physics Letters, vol. 124, no. 4, Jan. 2024. [Online]. Available: http://dx.doi.org/10.1063/5.0189542

    2024 doi

  23. [31]

    Macromagnetic simulation for reservoir computing utilizing spin dynamics in magnetic tunnel junctions,

    T. Furuta, K. Fujii, K. Nakajima, S. Tsunegi, H. Kubota, Y . Suzuki, and S. Miwa, “Macromagnetic simulation for reservoir computing utilizing spin dynamics in magnetic tunnel junctions,”Physical Review Applied, vol. 10, no. 3, Sep. 2018. [Online]. Available: http://dx.doi.org/...

    2018 doi

  24. [32]

    At the edge of chaos: Real-time computations and self-organized criticality in recurrent neural networks,

    N. Bertschinger, T. Natschl ¨ager, and R. Legenstein, “At the edge of chaos: Real-time computations and self-organized criticality in recurrent neural networks,”Advances in neural information processing systems, vol. 17, 2004

    2004

  25. [2014]

    Available: https://arxiv.org/abs/1412.6980

    [Online]. Available: https://arxiv.org/abs/1412.6980

    arXiv

  26. [2016]

    Available: http://dx.doi.org/10.18653/v1/W16-3210

    [Online]. Available: http://dx.doi.org/10.18653/v1/W16-3210

    doi

  27. [2024]

    Available: http://dx.doi.org/10.1063/5.0184848

    [Online]. Available: http://dx.doi.org/10.1063/5.0184848

    doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.