{"id":"e57a144d-0445-48b8-a7d1-3968c100893a","arxiv_id":"2505.02103","paper_version":5,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Equilibrium Propagation trains a simulated oscillator Ising machine to classify MNIST at 97.2% and Fashion-MNIST at 88.0%, with robustness to quantized parameters and phase noise.","lead":"An oscillator Ising machine, a chip built for combinatorial optimization, can also be trained to classify images using an energy-based learning rule called Equilibrium Propagation. Simulations reach about 97 percent accuracy on MNIST and stay accurate under realistic hardware imperfections such as coarse phase readout and quantization.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (11) requires negative synchronization fields S_y = -β/2 for the positive nudging phase, yet every result is a simulation with unconstrained S and no evidence is given that existing OIM hardware can program S with either sign.","rationale":"The reader identified the same weakest assumption: the unverified hardware availability of negative synchronization fields and sign-changing nudging. I agree that this is the most load-bearing concern because the paper's practical significance rests on retrofitting existing OIM hardware with no modifications. The mathematical mapping from the MSE loss to Eqs. (8)-(11) is correct, and the simulations appear self-consistent, so the concern is not about internal inconsistency but about an unverified physical realizability assumption. The simulation accuracy numbers, the EP-BPTT correspondence, and the quantization/noise studies are all conditional on Eq. (1) being executable with arbitrary S. If the sign is constrained, the positive nudging phase in the proposed cosine-output formulation cannot implement the intended loss. The concrete test would settle whether this is a real obstacle or a trivial programming detail; if the hardware can program S with either sign, the concern is resolved and the paper's claims stand. Other issues, such as the use of MNIST/100 for the hardware-robustness studies and the lack of released code, are secondary to this central realizability question. The reader's CONDITIONAL verdict is appropriate: the paper should be accepted only if the sign/range capability of S in existing OIM hardware is demonstrated or explicitly shown to be a configurable parameter.","tokens_in":7769,"tokens_out":15190,"duration_ms":209344,"concrete_test":"Rerun the MNIST/100 experiment of Section V with Eq. (11) modified to the hardware constraint S_y∈[0,S_max], preserving all other hyperparameters; if the positive-β phase cannot be set up because Eq. (7) requires S_y=-β/2, or if no alternative output readout such as y=cos(φ-α) recovers the reported 90.6±1.7% accuracy under S_y≥0, then the central claim of training on unmodified hardware is contradicted. A complementary check is to inspect a concrete CMOS OIM circuit (e.g., Ref. [19] or [18]) and determine whether the second-harmonic injection coefficient S is programmable negative in the first place.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's core promise is that an existing OIM can be repurposed for EP through parameter configuration alone. The mapping that makes this work is Eqs. (8)-(11), where the MSE loss is folded into the output bias h_y = b_y + βyhat and the synchronization field S_y = -β/2. For the positive nudging phase (β>0), this requires S_y<0. In the standard OIM model cited in Eq. (3), positive S is what stabilizes the binary phase states φ∈{0,π}; negative S inverts the second-harmonic potential, and the paper does not show that any physical OIM—CMOS VCO or otherwise—can deliver S<0. The MNIST and Fashion-MNIST numbers all come from Euler integration of Eq. (1) with S treated as an unconstrained real parameter. If real OIM hardware has a sign or range restriction on S, then the positive-β phase either cannot be realized or implements a different loss, and the 'no hardware modifications' claim fails. This is not a question of tuning: Eq. (7) fixes the coefficient of cos(2φ) to +1/4, which fixes the sign of S. The negative phase requires the opposite sign, so the hardware must support both signs, but the paper does not address this, despite asserting in the Discussion that current OIMs already contain all essential components.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":8073,"tokens_out":7913,"duration_ms":86143,"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":[{"comment":"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.","section":"Section IV, Eq. (11)"},{"comment":"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.","section":"Section V, Fig. 4"},{"comment":"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.","section":"Section V, Fig. 5"},{"comment":"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.","section":"Section V, Figs. 1 and 2"}],"minor_comments":[{"comment":"No code or data availability statement is included; the simulations are described as implemented in PyTorch but not released. Please add a reproducibility statement.","section":"General"},{"comment":"The text says phases are quantized to '2n evenly-spaced values'; this should be '2^n evenly-spaced values.'","section":"Sections III and V"},{"comment":"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.","section":"Section V"},{"comment":"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.","section":"Section V, Fig. 1"},{"comment":"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.","section":"Section VI, Discussion"}],"recommendation":"major_revision","confidential_remarks":"The core simulation results appear internally sound, but the hardware-compatibility claim is the main risk. The negative-S issue is concrete: if the authors can point to a circuit that programs synchronization fields with both signs, this would address it; otherwise the claims in the abstract and Discussion need to be scaled back to simulation-level feasibility. The quantization and noise experiments also need to be reported with less optimistic calibration assumptions before the hardware claims are accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The key question on this paper is whether an existing OIM can actually run the EP update as described. The mathematics checks out: folding the MSE loss into the output bias and the S_y = -β/2 term is a neat, internally consistent mapping, and the EP-BPTT correspondence in Fig. 1 is exactly the right sanity test. The simulation results on MNIST and Fashion-MNIST are plausible and in line with other EP implementations. The robustness experiments on phase-measurement quantization and parameter quantization are well designed and informative. As far as I can tell, the specific S_y mapping is not in the cited prior work, so that is a real new contribution.\n\nThe main soft spot is the hardware claim. For β>0 the mapping requires S_y<0. In the standard OIM energy, positive S creates the binary phase minima; negative S inverts that potential and destabilizes the 0/π states. The paper never shows that existing CMOS or VCO OIMs can program S with either sign, or that negative S is physically available without circuit changes. All simulations treat S as an unconstrained real parameter. The Discussion's assertion that 'current OIM implementations already incorporate all essential components' is therefore not supported by the evidence in the paper. That doesn't sink the simulation study, but it does mean the headline claim of 'no significant hardware modifications' is premature.\n\nOther issues are minor. The noise-benefit result re-optimizes β for each noise level, but the protocol for choosing β is not described; if that tuning used the test set, the conclusion is leaky. The quantization experiments calibrate the parameter ranges from full-precision runs, which also gives the scheme more information than a real deployment would have. There is no code, which is a real omission for a paper whose contribution is a concrete algorithmic mapping. And the 'first comprehensive framework' phrasing overstates what was already done by Wang et al. and Rageau & Grollier; the novelty is narrower, but it is real.\n\nWho should read this: people working on EP hardware, oscillator-based Ising machines, or on-chip learning. They will get a clean simulation-level demonstration that EP can be mapped onto OIM dynamics, and a useful robustness benchmark. I would cite it for the S_y mapping and the quantization results, with a caveat about the hardware sign issue. It deserves peer review; a referee should ask the authors to clarify negative S on real hardware, describe the β tuning loop, and provide code or at least detailed pseudocode.","headline":"Solid simulation study with a clean new mapping, but the 'no hardware modifications' claim relies on an untested assumption about negative synchronization fields.","tokens_in":8597,"tokens_out":4060,"would_cite":true,"duration_ms":47563,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Oscillator Ising machines can be trained by Equilibrium Propagation, reaching 97.2% test accuracy on MNIST.","keywords":["Oscillator Ising Machine","Equilibrium Propagation","neuromorphic computing","on-chip learning","coupled oscillators","MNIST classification","phase noise tolerance","energy-based learning"],"falsifier":"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.","tokens_in":7537,"feed_emoji":"🧠","tokens_out":6070,"duration_ms":58510,"temperature":0.7,"pith_summary":"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.","feed_headline":"Oscillator Ising machines train themselves, hitting 97.2% on MNIST","feed_subtitle":"Equilibrium Propagation turns optimization hardware into a learner, with no new circuitry.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the energy-gradient-descent property of OIMs that EP relies on.","marker":"[1]"},{"why":"Provides the prior Ising-machine training result on MNIST/100 that the paper's result is compared against.","marker":"[2]"},{"why":"Earlier attempt at training oscillator networks with EP, whose initialization and synchronization difficulties this paper addresses.","marker":"[3]"},{"why":"Supplies VCO calibration and microsecond timescale estimates used for the speedup projection.","marker":"[8]"},{"why":"Introduces Equilibrium Propagation and its energy-based update rule.","marker":"[9]"},{"why":"Establishes the symmetric positive/negative nudging gradient estimator used in the paper's update equations.","marker":"[10]"},{"why":"Defines Backpropagation Through Time, the reference for the gradient-correspondence validation.","marker":"[11]"},{"why":"Provides the MNIST dataset used for the main accuracy results.","marker":"[13]"},{"why":"Provides the Fashion-MNIST dataset used for the second accuracy result.","marker":"[14]"},{"why":"Conventional EP accuracy benchmark on full MNIST that the paper compares against.","marker":"[15]"}],"fun_headline_variants":["Oscillator Ising machines learn on-chip with no extra circuitry","Equilibrium Propagation trains oscillator Ising machines directly","Self-training oscillator Ising machines hit 97% on MNIST","Ising hardware gains on-chip learning, no new parts needed"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Oscillator Ising machines learn on-chip with no extra circuitry","Equilibrium Propagation trains oscillator Ising machines directly","Self-training oscillator Ising machines hit 97% on MNIST","Ising hardware gains on-chip learning, no new parts needed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000215,"raw_usage":{"total_tokens":1407,"prompt_tokens":902,"completion_tokens":505,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":435}},"tokens_in":518,"tokens_out":505,"duration_ms":5668,"temperature":1.0,"reasoning_tokens":435,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T01:02:56.339926+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"A note on analyzing the stability of oscillator Ising machines,","cited_arxiv_id":null,"evidence_quote":"Supplies the energy-gradient-descent property of OIMs that EP relies on."},{"cited_title":"Training an Ising machine with equilibrium propagation,","cited_arxiv_id":null,"evidence_quote":"Provides the prior Ising-machine training result on MNIST/100 that the paper's result is compared against."},{"cited_title":"Backpropagation through time: what it does and how to do it,","cited_arxiv_id":null,"evidence_quote":"Defines Backpropagation Through Time, the reference for the gradient-correspondence validation."},{"cited_title":"Gradient-based learning applied to document recognition,","cited_arxiv_id":null,"evidence_quote":"Provides the MNIST dataset used for the main accuracy results."},{"cited_title":"Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms,","cited_arxiv_id":null,"evidence_quote":"Provides the Fashion-MNIST dataset used for the second accuracy result."}],"review_version":1}