Pith. sign in

REVIEW 4 major objections 5 minor 20 references

How to Train an Oscillator Ising Machine using Equilibrium Propagation

T0 review · 4 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read Oscillator Ising machines can be trained by Equilibrium Propagation, reaching 97.2% test accuracy on MNIST.

desk verdict Solid simulation study with a clean new mapping, but the 'no hardware modifications' claim relies on an untested assumption about negative synchronization fields. read the letter →

arxiv 2505.02103 v5 pith:7ALYVULC submitted 2025-05-04 cond-mat.dis-nn

Alex Gower This is my paper · ORCID
classification cond-mat.dis-nn
keywords OscillatorIsingMachineEquilibriumPropagationneuromorphiccomputingon-chiplearningcoupledoscillatorsMNISTclassificationphasenoisetoleranceenergy-based
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

This paper argues that an Oscillator Ising Machine, a physical network of coupled oscillators whose phases descend an energy landscape, can be turned into a trainable neural network using Equilibrium Propagation, with no hardware changes beyond parameter settings. The author shows in simulation that mapping the training objective onto the OIM's energy function produces local, hardware-compatible update rules, and that these rules train a one-hidden-layer network to about 97.2% test accuracy on MNIST and 88.0% on Fashion-MNIST. The practical stake is that devices currently built for combinatorial optimization could be repurposed as fast, energy-efficient neuromorphic learners, with robustness to realistic constraints such as 10-bit parameter precision, 4-bit phase readout, and phase noise.

What carries the argument

The carrying object is the OIM energy function $V=-\frac12\sum_{i\ne j} J_{ij}\cos(\phi_i-\phi_j)-\sum_i h_i\cos\phi_i-\sum_i \frac{S_i}{2}\cos(2\phi_i)$, together with the EP total energy $F=E+\beta\ell$. The mechanics is a term-by-term mapping of network weights, biases, and the MSE loss onto $J_{ij}$, $h_i$, and $S_i$, so that the physical phase dynamics are simultaneously the free-phase and nudged-phase relaxations of EP. The update rule $\Delta\theta=\eta \hat\nabla_{\rm EP}(\beta)$ reduces to local products and differences of $\cos$ of phases, meaning each synapse can compute its own gradient from signals present at its two endpoints.

What would settle it

If, on an actual chip, the output-synchronization strength cannot be set negative without destabilizing the oscillators, or if the measured training updates do not match the simulated ones beyond the reported tolerance, the central claim fails. A direct experiment would program the parameter mapping on a fabricated CMOS oscillator array and compare the free and nudged phase fixed points with the simulated ones.

Watch

Extended reading notes

Core claim

The central claim is that the energy-gradient-descent dynamics of an OIM are exactly the substrate Equilibrium Propagation needs: by identifying the OIM energy with the EP total energy $F=E+\beta\ell$, the author derives parameter mappings under which free and nudged phases implement the EP gradient estimator. The key identity is that the mean-squared-error loss $\frac12\sum_i(\cos\phi_i^{(y)}-\hat y_i)^2$ expands into oscillator-compatible terms, fixing the output synchronization field at $S_i^{(y)}=-\beta/2$. With this mapping, the parameter updates are local functions of directly connected oscillator phases. Simulated training reaches $97.2\pm0.1\%$ on MNIST and $88.0\pm0.1\%$ on Fashion-MNIST, and the updates match Backpropagation Through Time gradients, confirming that the gradient-descending property holds despite the trigonometric coupling.

Load-bearing premise

The load-bearing premise is that real oscillator chips can be configured to apply the negative output-synchronization strengths and the sign-flipping nudging forces the training rule requires; the paper's evidence for this is simulation, not a hardware measurement.

Editorial extensions

If this is right

  • Existing OIM hardware designed for combinatorial optimization can be repurposed as neuromorphic processors by software-level parameter configuration, since all required energy terms already exist in its energy function.
  • OIM-based EP training matches BPTT gradients on the tested architectures, so accuracy on these tasks is limited by the network architecture rather than by the training algorithm.
  • Training remains accurate with 10-bit quantized parameters and 4-bit phase readout, meaning the required analog and digital precision is within current CMOS capabilities.
  • With the nudging factor tuned so that $\beta\gtrsim \xi/2$, moderate phase noise does not degrade accuracy and can slightly help, reducing the need for aggressive noise suppression.
  • On a GHz-frequency physical implementation, the same MNIST training could run in seconds to minutes rather than tens of hours, a several-orders-of-magnitude speedup over simulation.

Reading between the lines

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

  • If a physical CMOS oscillator array can realize negative synchronization fields, the same energy-mapping recipe should extend to deeper and convolutional oscillator networks, since the locality of the updates does not depend on network depth.
  • The observed benefit of noise with $\beta\gtrsim\xi/2$ suggests a testable design rule: deliberately inject phase noise during training as a regularizer, rather than spending area on noise suppression.
  • The parameter mapping may transfer to other physical systems that minimize a trigonometric energy, such as coupled phase-oscillator arrays, provided their couplings can be set per pair and their outputs read as $\cos\phi$.
  • A decisive next experiment, not performed in the paper, is to program the parameter mapping on a fabricated OIM and check that the free and nudged fixed points and the resulting updates match simulation within the claimed noise tolerance.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 manuscript proposes that Oscillator Ising Machines (OIMs) can be trained on-chip via Equilibrium Propagation (EP) by mapping the MSE loss into the OIM energy function through Eqs. (7)-(11), deriving local EP update rules (Eqs. (12)-(15)), and validating the approach in simulations on MNIST, Fashion-MNIST, and MNIST/100. The simulations report about 97.2% test accuracy on full MNIST, 88.0% on Fashion-MNIST, and 90.6% on MNIST/100, with additional experiments on phase-measurement quantization, parameter quantization, and phase noise. The central claim is that this works 'without requiring any significant hardware modifications' to existing CMOS OIM designs.

Significance. If the central claim holds, this would be a meaningful step toward repurposing combinatorial-optimization hardware as neuromorphic learners, and the paper's main conceptual contribution is the explicit mapping from EP to the OIM energy function. The mathematical mapping in Eqs. (7)-(11) is internally consistent, the EP-BPTT correspondence test in Fig. 1 is the right sanity check for the gradient-descending update property, and the update rules are local and therefore plausible for on-chip learning. The accuracy results are competitive with prior EP and Ising-hardware implementations, and the quantization and noise studies address practically relevant hardware constraints. However, the hardware-compatibility claim depends on assumptions about sign control and calibration that are not supported by the presented evidence, and the robustness experiments contain optimistic calibration choices. The manuscript does not provide code or data, which limits reproducibility.

major comments (4)
  1. [Section IV, Eq. (11)] The implementation of the MSE loss in Eq. (7) fixes the coefficient of cos(2φ_i) to +β/4 in the total energy, which through Eq. (3) requires S_i^{(y)} = -β/2. For the positive nudging phase (β>0) this requires a negative synchronization field, i.e., the second-harmonic potential is inverted and the output phases are not stabilized toward φ∈{0,π} but toward quadrature phases. The paper does not provide any circuit-level evidence or reference showing that existing CMOS OIM implementations (Refs. [8], [18], [19]) can program S with either sign or can switch the sign of β in the control electronics, yet the abstract and Discussion claim that EP can be implemented 'without requiring any significant hardware modifications.' This is a load-bearing gap: if the hardware cannot realize S<0, the positive-nudging phase implements a different loss and the central claim fails. Please either supply hardware evidence or explicitly restrict the claim to simulation-level feasibility.
  2. [Section V, Fig. 4] In the parameter-quantization experiment, parameters are rounded to 2^n evenly spaced values 'within their typical training ranges (determined from full-precision runs)' before each dynamics phase. This assumes oracle knowledge of the full-precision ranges for every parameter at every phase; in a real training run those ranges are not known in advance, and any online range calibration would itself be subject to quantization errors. As written, the 10-bit robustness result therefore does not directly support the 'viable for physical implementations' claim. Please state explicitly how and when the ranges are computed, and test a fixed or online-calibrated range scheme.
  3. [Section V, Fig. 5] The noise-tolerance conclusion in Section V and the abstract is based on re-optimizing β for each noise level: at ξ=0.2, accuracy is 92.0±0.3% with β=0.3, while with the noise-free-optimized β=0.05 accuracy falls to 41.0±6.2%. This demonstrates that noise can be compensated by retuning, not that noise is intrinsically beneficial. Please tone down the 'noise can potentially be beneficial' wording and report results with either a fixed β or a concrete β-scheduling rule that would be available in practice.
  4. [Section V, Figs. 1 and 2] Equilibrium Propagation's update rule (4) is defined at stationary points, and the text says the free and nudged dynamics are run 'until convergence,' but no convergence criterion, residual threshold, or check for multistability is reported. Because Eq. (1) has a trigonometric energy landscape with multiple fixed points and possible limit cycles, the accuracy numbers in Fig. 2 and the EP-BPTT comparison in Fig. 1 are not verifiably tied to EP's fixed-point assumption. Please report convergence metrics (e.g., final |dφ/dt|, or a stability check) or otherwise justify that all simulated phases reach fixed points.
minor comments (5)
  1. [General] No code or data availability statement is included; the simulations are described as implemented in PyTorch but not released. Please add a reproducibility statement.
  2. [Sections III and V] The text says phases are quantized to '2n evenly-spaced values'; this should be '2^n evenly-spaced values.'
  3. [Section V] The noise notation 'ξN(0,1)' is ambiguous: ξ appears to denote a noise amplitude multiplied by a standard normal draw, but it is written as if ξ were the random variable. Please define the noise term explicitly, e.g., dφ_i ← dφ_i + ξ dW_i with ξ a scalar amplitude.
  4. [Section V, Fig. 1] In Fig. 1, the network size, dataset, and hyperparameters used for the EP-BPTT comparison are not stated; please specify them in the caption or text.
  5. [Section VI, Discussion] The speedup estimate (seconds to minutes vs. 40-60 hours) relies on unstated assumptions about oscillator convergence times and measurement overhead; please provide a reference or a detailed calculation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the parameter mapping is an algebraic embedding of EP into the OIM energy, and accuracy is measured against external benchmarks.

full rationale

The paper's core derivation is self-contained: Eqs. (8)-(11) are an explicit construction showing that the total energy F = E + beta*ell for an MSE loss coincides, term by term, with the OIM energy (3), so the OIM dynamics implement EP. The parameter updates (12)-(15) follow by differentiating this energy, not by fitting the reported accuracies. MNIST, Fashion-MNIST, and MNIST/100 are external benchmarks, and the reported 97.2%, 88.0%, and 90.6% figures are simulation outcomes rather than enforced targets. The EP-BPTT correspondence in Fig. 1 is obtained by moving into the theoretically expected small-beta, long-free-phase regime; the paper states this tuning transparently, and it is a consistency check rather than a load-bearing prediction. There are no self-citations used as premises, no imported uniqueness theorem, and no fitted parameter renamed as a prediction. The negative-synchronization-field issue raised by Eq. (11) is a hardware-realizability concern, not a circularity: it questions whether a physical OIM can supply S < 0, but it does not make the derivation depend on its own conclusion.

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

The central result is a simulation built from published EP theory and OIM dynamics; it introduces no new physical entity and few free parameters beyond ordinary training hyperparameters. The most load-bearing assumptions are hardware capability (negative S and beta) and fixed-point convergence, both unverified.

free parameters (4)
  • Nudging factor beta = 0.1 (MNIST/FMNIST), 0.05 (MNIST/100 noise-free), 0.3 (MNIST/100 with xi=0.2)
    Chosen per dataset and noise level; controls EP gradient bias. The paper itself notes beta >= xi/2 is needed empirically, not derived.
  • Integration step size epsilon and phase durations T, K = epsilon=0.45 or 0.5, T=4000 or 3500, K=400 or 350
    Chosen for stable convergence; no discretization error analysis is reported.
  • Learning rates, batch size, epochs = eta_w^h=0.01, eta_w^y=0.001, eta_b=0.001, batch=128 or 20, epochs=50
    Standard hyperparameters selected empirically; listed in Table I.
  • Quantization range calibration = Ranges determined from full-precision training runs
    In quantization experiments, the 2^n grid is set using full-precision training ranges, which is oracle information not necessarily available at hardware deployment.
assumptions (5)
  • domain assumption Oscillator dynamics (1) are equivalent to gradient descent on energy V (3)
    Taken from Bashar et al. [1]; foundation of all EP updates.
  • domain assumption The system reaches a stable fixed point for every configuration used, and the free-phase fixed point is in the basin needed for nudged phases
    No fixed-point uniqueness or multistability analysis is given; EP gradient estimates require settled stationary states.
  • ad hoc to paper Physical OIM hardware can set negative synchronization fields S_y=-beta/2 and sign-changing beta for positive and negative nudging
    Required by Eqs. (10)-(11); the 'no hardware modifications' claim depends on this but it is not demonstrated.
  • standard math EP update identities (4)-(6) apply at finite beta and finite phase durations
    Taken from Scellier and Bengio and Laborieux et al.; they are exact only in the beta-to-0 limit, while experiments use beta=0.05 to 0.3.
  • domain assumption Euler integration with step epsilon=0.45 to 0.5 faithfully represents the continuous OIM dynamics
    No convergence test or comparison with smaller step sizes is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of How to Train an Oscillator Ising Machine using Equilibrium Propagation." pith.science (2026). https://pith.science/paper/7ALYVULC

@misc{pith2026250502103,
  author       = {Pith},
  title        = {Pith review of: How to Train an Oscillator Ising Machine using Equilibrium Propagation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7ALYVULC}},
  note         = {Machine review of arXiv:2505.02103}
}
abstract

We show that Oscillator Ising Machines (OIMs) are prime candidates for use as neuromorphic machine learning processors with Equilibrium Propagation (EP) based on-chip learning. The inherent energy gradient descent dynamics of OIMs, combined with their standard CMOS implementation using existing fabrication processes, provide a natural substrate for EP learning. Our simulations confirm that OIMs satisfy the gradient-descending update property necessary for a scalable Equilibrium Propagation implementation and achieve $\sim 97.2\pm0.1\%$ test accuracy on MNIST and $\sim 88.0\pm0.1\%$ on Fashion-MNIST without requiring any significant hardware modifications. Importantly, OIMs maintain robust performance under realistic hardware constraints, including 10-bit parameter quantization, 4-bit phase measurement precision, and moderate phase noise that can potentially be beneficial with parameter optimization. These results establish OIMs as a promising platform for fast and energy-efficient neuromorphic computing, potentially enabling energy-based learning algorithms that have been previously constrained by computational limitations.

Figures

Figures reproduced from arXiv: 2505.02103 by the authors.

Figure 1
Figure 1. EP-BPTT correspondence showing symmetric Equilibrium Prop [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Test accuracy comparison between OIM architectures on MNIST [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Parameter quantization effects: Test accuracy vs. training epochs at [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Phase noise effects on training with fixed vs. optimized parameters for [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 11 canonical work pages

  1. [8]

    Solving combinatorial optimisation problems using oscillator based Ising machines,

    T. Wang, L. Wu, P. Nobel, and J. Roychowdhury, “Solving combinatorial optimisation problems using oscillator based Ising machines,” Natural Computing, vol. 20, no. 2, pp. 287–306, Jun. 2021

    2021

  2. [18]

    Experiments with an oscillator based Ising machine,

    S. Roy and B. Ulmann, “Experiments with an oscillator based Ising machine,” Feb. 2025, arXiv:2502.03167 [cs] version: 1. [Online]. Available: http://arxiv.org/abs/2502.03167

    arXiv 2025

  3. [19]

    An integrated coupled oscillator network to solve optimization problems,

    M. Graber and K. Hofmann, “An integrated coupled oscillator network to solve optimization problems,” Communications Engineering , vol. 3, no. 1, p. 116, Aug. 2024, publisher: Nature Publishing Group. [Online]. Available: https://www.nature.com/articles/s44172-024-00261-w

    work page 2024

  4. [1]

    A note on analyzing the stability of oscillator Ising machines,

    M. K. Bashar, Z. Lin, and N. Shukla, “A note on analyzing the stability of oscillator Ising machines,” Electronics Letters , vol. 59, no. 24, p. e13054, 2023, eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1049/ell2.13054

    work page doi:10.1049/ell2.13054 2023

  5. [2]

    Training an Ising machine with equilibrium propagation,

    J. Laydevant, D. Markovi ´c, and J. Grollier, “Training an Ising machine with equilibrium propagation,” Nature Communications, vol. 15, no. 1, p. 3671, Apr. 2024, publisher: Nature Publishing Group. [Online]. Available: https://www.nature.com/articles/s41467-024-46879-4

    work page 2024

  6. [3]

    Training and synchronizing oscillator networks with Equilibrium Propagation,

    T. Rageau and J. Grollier, “Training and synchronizing oscillator networks with Equilibrium Propagation,” Neuromorphic Computing and Engineering , vol. 5, no. 3, p. 034008, Jul. 2025, publisher: IOP Publishing. [Online]. Available: https://dx.doi.org/10.1088/2634- 4386/adebaa

    doi:10.1088/2634- 2025

  7. [4]

    Estimation of energy consumption in machine learning,

    E. Garc ´ıa-Mart´ın, C. F. Rodrigues, G. Riley, and H. Grahn, “Estimation of energy consumption in machine learning,” Journal of Parallel and Distributed Computing, vol. 134, pp. 75–88, Dec. 2019. [Online]. Avail- able: https://linkinghub.elsevier.com/retrieve/pii/S0743731518308773

    work page 2019

  8. [5]

    Coupled oscillators for computing: A review and perspective,

    G. Csaba and W. Porod, “Coupled oscillators for computing: A review and perspective,” Applied Physics Reviews , vol. 7, no. 1, p. 011302, Jan. 2020. [Online]. Available: https://doi.org/10.1063/1.5120412

    work page doi:10.1063/1.5120412 2020

Show all 20 references
  1. [6]

    Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture,

    M. Assran, Q. Duval, I. Misra, P. Bojanowski, P. Vincent, M. Rabbat, Y . LeCun, and N. Ballas, “Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture,” Apr. 2023, arXiv:2301.08243 [cs]. [Online]. Available: http://arxiv.org/abs/2301.08243

    2023 arXiv

  2. [7]

    Training Coupled Phase Oscillators as a Neuromorphic Platform using Equilibrium Propagation,

    Q. Wang, C. C. Wanjura, and F. Marquardt, “Training Coupled Phase Oscillators as a Neuromorphic Platform using Equilibrium Propagation,” Feb. 2024, arXiv:2402.08579 [cond-mat, physics:physics]. [Online]. Available: http://arxiv.org/abs/2402.08579

    2024 arXiv

  3. [9]

    Equilibrium Propagation: Bridging the Gap Between Energy-Based Models and Backpropagation,

    B. Scellier and Y . Bengio, “Equilibrium Propagation: Bridging the Gap Between Energy-Based Models and Backpropagation,” Mar. 2017, arXiv:1602.05179. [Online]. Available: http://arxiv.org/abs/1602.05179

    2017 arXiv

  4. [10]

    Scaling Equilibrium Propagation to Deep ConvNets by Drastically Reducing its Gradient Estimator Bias,

    A. Laborieux, M. Ernoult, B. Scellier, Y . Bengio, J. Grollier, and D. Querlioz, “Scaling Equilibrium Propagation to Deep ConvNets by Drastically Reducing its Gradient Estimator Bias,” Jun. 2020, arXiv:2006.03824. [Online]. Available: http://arxiv.org/abs/2006.03824

    2020 arXiv

  5. [11]

    Backpropagation through time: what it does and how to do it,

    P. Werbos, “Backpropagation through time: what it does and how to do it,” Proceedings of the IEEE , vol. 78, no. 10, pp. 1550–1560, Oct

  6. [12]

    Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification,

    K. He, X. Zhang, S. Ren, and J. Sun, “Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification,” Feb. 2015, arXiv:1502.01852 [cs]. [Online]. Available: http://arxiv.org/abs/1502.01852

    2015 arXiv

  7. [13]

    Gradient-based learning applied to document recognition,

    Y . Lecun, L. Bottou, Y . Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,” Proceedings of the IEEE , vol. 86, no. 11, pp. 2278–2324, Nov. 1998. [Online]. Available: https://ieeexplore.ieee.org/document/726791

    1998

  8. [14]

    Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms,

    H. Xiao, K. Rasul, and R. V ollgraf, “Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms,” Aug. 2017, aDS Bibcode: 2017arXiv170807747X. [Online]. Available: https://ui.adsabs.harvard.edu/abs/2017arXiv170807747X

    2017

  9. [15]

    Updates of Equilibrium Prop Match Gradients of Backprop Through Time in an RNN with Static Input,

    M. Ernoult, J. Grollier, D. Querlioz, Y . Bengio, and B. Scellier, “Updates of Equilibrium Prop Match Gradients of Backprop Through Time in an RNN with Static Input,” May 2019, arXiv:1905.13633 [cs]. [Online]. Available: http://arxiv.org/abs/1905.13633

    2019 arXiv

  10. [16]

    Training Deep Boltzmann Networks with Sparse Ising Machines,

    S. Niazi, N. A. Aadit, M. Mohseni, S. Chowdhury, Y . Qin, and K. Y . Camsari, “Training Deep Boltzmann Networks with Sparse Ising Machines,” Nature Electronics , vol. 7, no. 7, pp. 610–619, Jun. 2024, arXiv:2303.10728 [cs]. [Online]. Available: http://arxiv.org/abs/2303.10728

    2024 arXiv

  11. [17]

    Temporal Spike Sequence Learning via Backpropagation for Deep Spiking Neural Networks,

    W. Zhang and P. Li, “Temporal Spike Sequence Learning via Backpropagation for Deep Spiking Neural Networks,” in Advances in Neural Information Processing Systems , vol. 33. Curran Associates, Inc., 2020, pp. 12 022–12 033. [Online]. Available: https://proceedings.neurips.cc/pa...

    2020

  12. [1990]

    Available: https://ieeexplore.ieee.org/document/58337

    [Online]. Available: https://ieeexplore.ieee.org/document/58337

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.