{"id":"02feab3b-a7ea-42c7-b753-1d461dbd4e35","arxiv_id":"2412.12783","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Using measured noise from a stochastic magnetic tunnel junction as perturbations, the DANP learning rule trains multilayer networks in simulation with accuracy approaching backpropagation.","lead":"This paper tests whether a local learning rule called node perturbation can train neural networks using the random switching noise of magnetic tunnel junctions, rather than clean Gaussian noise. Simulations show it can approach backpropagation accuracy, and a small proof-of-concept uses a real device as the noise source.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The load-bearing assumption is that the proposed analog circuit in §II.C reproduces the DANP dynamics simulated in Python; the §III.D hardware demo only feeds Arduino-sampled sMTJ bits into a Python network.","rationale":"The reader's weakest assumption correctly identifies the gap: the analog circuit is proposed but never implemented, and the experimental validation substitutes a Python network for it. My stress-test confirms this is the single most load-bearing concern because the abstract's flagship sentence explicitly asserts learning in a network composed of physical sMTJs. The multilayer simulations with measured sMTJ time series are a meaningful positive result for the DANP algorithm, and the paper is transparent in §III.D about the Arduino/Python split; however, transparency does not close the gap between the demonstrated system and the claimed physical implementation. I agree with the CONDITIONAL verdict: the algorithmic result is credible, but the hardware claim should not be accepted as stated without a test of the analog circuit or an explicit narrowing of the claim. Hyperparameter selection on test accuracy (Appendix A) is a secondary weakness, not the main load-bearing issue, since it does not affect the central mismatch between the claimed physical demonstration and what was actually run.","tokens_in":11257,"tokens_out":3247,"duration_ms":33896,"concrete_test":"Build or faithfully hardware-emulate the Figure 1 analog network: fixed resistor weights, op-amp summing amplifiers, diode-based ReLU, and sMTJ noise injected at each node. Train the same 2×2×2×1 ground-truth task for 100 epochs as in §III.D. If the loss decreases comparably to Figure 5, the physical-learning claim is supported. A cheaper intermediate check: measure the input/output transfer function of one op-amp-plus-diode stage with actual sMTJ noise and compare it with the simulated activation and noise model over the training input distribution; quantify the maximum deviation and rerun the Figure 4 simulations with that measured distortion to see whether convergence survives.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract claims 'learning can be achieved in a small network composed of physical stochastic magnetic tunnel junctions,' but the only physical hardware in §III.D is an Arduino reading a single sMTJ. The 2×2×2×1 network, forward pass, loss, and weight updates all run in Python, and the noise values are shuffled from a 200-sample buffer before being injected. The proposed analog implementation of Figure 1—op-amp summing amplifier, diode-based ReLU, resistive weights, and per-node sMTJ noise injection—is never built. Thus the central claim rests on an untested equivalence: that real sMTJ noise fed into that analog circuit would produce the same learning dynamics as the simulated DANP update of Eq. 4. Unmodeled effects such as sMTJ two-level quantization interacting with the summing amplifier, finite output impedance, diode nonlinearity and leakage, resistor weight mismatch, and per-node noise bandwidth could change the effective update enough to prevent convergence. This is not a fatal flaw in the algorithm: the multilayer simulations in Figure 4 using raw sMTJ time series provide genuine evidence that autocorrelated two-level noise can train DANP networks. But those simulations do not test the device physics of the proposed circuit, so the hardware-level claim is stronger than the evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes using the intrinsic stochastic noise of superparamagnetic magnetic tunnel junctions (sMTJs) as the perturbation source for decorrelated activity-based node perturbation (DANP), a local learning rule that avoids backpropagation. After recapping DANP in Section II.A, the paper characterizes measured sMTJ noise, fits a two-state hidden Markov model to it, and proposes an analog circuit concept in Figure 1 in which sMTJ noise is injected at each node. Section III reports three levels of evidence: single-layer ANP robustness sweeps on MNIST (Figure 3); multilayer DANP simulations on MNIST, CIFAR-10, and CIFAR-100 using both simulated and measured sMTJ noise (Figure 4); and a proof-of-concept experiment in which an Arduino reads a single sMTJ and feeds its voltage samples to a 2x2x2x1 network that otherwise runs in Python (Figure 5). The central numerical claim is that DANP with real, autocorrelated, two-level sMTJ noise learns effectively in multilayer networks and approaches backpropagation performance. The abstract further claims that learning was demonstrated experimentally 'in a small network composed of physical stochastic magnetic tunnel junctions,' but the hardware experiment does not implement the network in physical sMTJ hardware.","tokens_in":11515,"tokens_out":4621,"duration_ms":48032,"significance":"If the claims are restricted to what is actually demonstrated, this is a valuable contribution: the simulations show that strongly autocorrelated two-level noise, including noise sampled sequentially from a single physical device, can drive DANP training in deep networks without requiring the idealized Gaussian perturbations of the original ANP formulation. The work also provides a useful benchmark for noise-based learning in physical substrates, and it ships code and data (Appendix C), which aids reproducibility. I find no circularity in the learning evaluation: the HMM parameters are fit to device measurements, and the algorithm is compared against backpropagation as an external baseline. The main weakness is the gap between the proposed analog learning circuit and the hardware experiment, which uses the sMTJ only as a noise source for a software network. With the abstract and hardware claims revised to match the evidence, the paper would be a sound and informative step toward noise-based physical learning.","major_comments":[{"comment":"The abstract's statement that learning can be achieved 'in a small network composed of physical stochastic magnetic tunnel junctions' is not supported by the experiments in §III.D. There, the network runs in Python: the Arduino reads a single sMTJ and supplies voltage samples, which are stored in a 200-sample buffer and shuffled before injection; the forward pass, loss, and weight updates all occur in software. The analog circuit in Figure 1 is a proposal described in §II.C, and Section IV explicitly defers embedding the forward pass, decorrelation, and parameter updates in physical hardware to future work. The load-bearing equivalence between Eq. (4) and the proposed analog circuit is therefore untested; unmodeled effects such as the sMTJ's two-level quantization interacting with the summing amplifier, diode nonlinearities, and resistive weight mismatch could alter the effective update. Please revise the claims to state that a single physical sMTJ was used as a noise source for an otherwise simulated DANP network, or provide a physical implementation of the network itself.","section":"Abstract; §III.D; §II.C"},{"comment":"The hardware proof-of-concept reports decreasing loss over 100 epochs, but Figure 5 shows no baseline condition, no chance-level reference, and no statistical test. Without a control, a reader cannot distinguish learning from a downward drift or from a favorable initialization. I recommend adding a control condition with fixed random noise or with the noise source disconnected, and reporting the distribution of final loss across multiple network initializations.","section":"§III.D; Fig. 5"},{"comment":"The hyperparameters for the multilayer simulations were selected by sweeping η and ϵ and choosing the pair with the highest test accuracy at 1000 epochs for a single seed. This test-set-based selection makes the reported test curves optimistic and complicates the comparison with backpropagation, because the same selection protocol is not described for the BP baselines. I suggest reporting performance under the same selection rule for BP, or performing selection on a held-out validation split and reporting test performance only for the final configuration.","section":"Appendix A"}],"minor_comments":[{"comment":"The loss is defined in Eq. (1) as squared error L = ||y* - y||^2, but Appendix A states that categorical cross-entropy was used for training. Please unify the notation and state the loss used in each experiment.","section":"§II.A; Appendix A"},{"comment":"The text states that the Arduino Uno has 14-bit resolution on its analog pins; the Arduino Uno's ADC is 10-bit. Please correct this factual detail or specify the exact hardware used.","section":"§III.D"},{"comment":"Reference [38] is listed as accessed on 2010-09-30, which predates the Telemetrix project; the access date is likely a typo and should be corrected.","section":"References"},{"comment":"The text says ANP is robust to 'Gaussian or Binomial distribution sampled noise,' but the panels of Figure 3 are not labeled clearly enough to identify which noise distribution each panel corresponds to. Please add explicit panel labels.","section":"§III.B; Fig. 3"},{"comment":"The sentence 'This is the case for a resistance-based setup, as shown here' is ambiguous; please clarify whether the series connection of sMTJs applies to the voltage-adder circuit in Figure 1 or to the multi-level noise generalization.","section":"§II.C"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's core algorithmic contribution is credible and the simulation evidence with real device noise is a genuine step forward. The main issue is a mismatch between the abstract's hardware claim and the actual experimental setup, which is a Python network with an sMTJ noise source. I believe this is fixable by rewriting the claims and possibly adding a control condition, rather than by new device fabrication. I would also ask the editor to ensure that the revised version clearly separates 'proposed circuit' from 'demonstrated hardware' in the title and abstract, as the current wording is likely to mislead readers outside the immediate field."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: the core simulation result is real and useful—DANP trains multilayer networks with measured, autocorrelated, two-level sMTJ noise and approaches backpropagation on MNIST and CIFAR. That is a genuine step beyond the idealized Gaussian noise in the authors' prior work. But the abstract overstates the hardware: the only physical component in the demo is the noise source; the network, forward pass, and weight updates run in Python.\n\nWhat's new: applying DANP to real sMTJ time series, both as a single measured source and via an HMM fit, and showing it works even in the worst-case serial injection. The Figure 4 simulations are solid: they benchmark against BP, include envelopes over seeds, and use both synthetic and measured noise. The Arduino proof-of-concept is a small but honest demonstration that live sMTJ noise can drive loss reduction in a toy network, though the network is simulated.\n\nSoft spots, in proportion: the abstract's phrase 'learning can be achieved in a small network composed of physical stochastic magnetic tunnel junctions' is not supported by Section III.D. The proposed analog circuit in Figure 1 is never built, so the equivalence between simulated DANP dynamics and the proposed op-amp/diode/resistive implementation is untested. That is a real gap, but it's a presentation problem rather than a fatal flaw in the algorithm. Also, the Appendix A hyperparameter sweeps select on test accuracy, which likely inflates reported performance; and code/data are only for Figure 4, not the Arduino demo. These are minor but worth fixing.\n\nWho it's for: people in neuromorphic hardware, spintronics, or alternative training rules will get value from the demonstration that real device noise can substitute for Gaussian in node perturbation. It deserves a serious referee, with the expectation that the authors either soften the hardware claim or back it with actual hardware. I'd send it to review and ask for a revised abstract and a note on the test-selection issue.","headline":"DANP learning genuinely works with real sMTJ noise in simulation, but the abstract's hardware claim outruns what was actually built: only the noise source is physical, not the network.","tokens_in":12106,"tokens_out":1744,"would_cite":true,"duration_ms":17736,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper demonstrates that the two-level, autocorrelated noise of a stochastic magnetic tunnel junction is sufficient to train multilayer neural networks under the decorrelated activity-based node perturbation rule, approaching…","keywords":["node perturbation","stochastic magnetic tunnel junctions","backpropagation-free learning","physical neural networks","noise-based learning","decorrelated activity-based node perturbation","superparamagnetism","local learning"],"falsifier":"If a full analog 2x2x2x1 network built from the proposed sMTJ noise-injection circuit is trained and its loss does not decrease over 100 epochs, or if the same loss decrease occurs when the sMTJ is replaced with a fixed voltage, then the claim that physical sMTJ noise drives learning is falsified.","tokens_in":11042,"feed_emoji":"🧲","tokens_out":7256,"duration_ms":68504,"temperature":0.7,"pith_summary":"This paper tries to show that the intrinsic noise of physical devices—specifically the spontaneous two-level resistance fluctuations of stochastic magnetic tunnel junctions—can be used as the learning signal for neural networks, instead of being filtered out. The authors build on the decorrelated activity-based node perturbation (DANP) algorithm, which needs only two forward passes under different noise perturbations and a scalar loss difference, so it avoids backpropagation and avoids measuring the noise itself. In simulation, DANP with real sMTJ noise approaches backpropagation accuracy on MNIST and CIFAR-10/100, even when a single strongly autocorrelated sMTJ time series supplies noise to all nodes. They also report a live proof of concept in which a physical sMTJ, read by a microcontroller, drives loss reduction in a small network running in software. If these results hold, learning in physical hardware would not require clean devices or gradient machinery; it could simply embrace whatever noise the hardware already produces.","feed_headline":"Magnetic junction noise alone trains neural networks","feed_subtitle":"A node-perturbation rule reaches near-backprop accuracy on MNIST and CIFAR using only perturbed forward passes.","key_machinery":"The central machinery is the decorrelated activity-based node perturbation (DANP) rule: after two forward passes with different noise perturbations, the update to layer $l$ weights is $\\delta W_l = N \\langle \\delta L \\, \\delta a_l / \\|\\delta a\\|^2 \\, \\bar{x}_{l-1}^\\top \\rangle$, where $\\delta L$ is the loss difference, $\\delta a$ the activity difference, and $\\bar{x}_{l-1}$ the decorrelated input, with the decorrelation transform $\\bar{x}_l = R_l x_l$ learned by an iterative local rule. Because it compares two noisy passes, DANP needs no clean reference pass, no access to the injected noise values, and no measurement of spatial correlations between nodes. The physical noise source is a stochastic magnetic tunnel junction, a magnetic device whose resistance randomly toggles between two levels via thermally driven magnetization switching.","core_discovery":"The paper's claim is that the two-level, autocorrelated, non-Gaussian noise actually produced by a stochastic magnetic tunnel junction is a sufficient learning signal for node-perturbation training. In simulations, multilayer networks with three hidden layers of 500 units trained by DANP with noise drawn either from a hidden Markov model fitted to a measured sMTJ time series or from that same time series fed serially to all nodes reach accuracies approaching those of backpropagation on MNIST, CIFAR-10, and CIFAR-100; the serially fed worst-case noise with strong autocorrelation still learns. A live setup, in which a microcontroller reads a single sMTJ and passes its samples to a Python 2x2x2x1 network, shows loss decreasing over 100 epochs. The paper also gives a circuit proposal for a fully analog version, while noting that the complete CMOS learning circuitry is left for future work.","pith_inferences":["A direct test the paper leaves implicit is whether the proposed analog circuit, with resistive weights, diode activation, and sMTJ voltage-adder noise injection, reproduces the simulated DANP dynamics; the reported experiment keeps the network itself in software.","Because the worst-case single-source autocorrelated noise still trains nearly as well as independent simulated sources, the effective requirement on the physical noise may be much weaker than white noise, broadening the class of usable devices.","Since DANP uses only a global scalar feedback signal, the same rule should transfer to physical systems whose forward dynamics are not differentiable, such as optical or mechanical networks, where backpropagation cannot be defined.","If sMTJ fluctuation rates can be tuned from milliseconds to nanoseconds, the approach makes a concrete energy-accuracy trade-off: faster noise speeds up learning iterations but raises the bandwidth and power cost of the readout and update circuitry."],"forward_implications":["Learning in a physical network requires only two noisy forward passes and a scalar loss difference; no backward pass and no clean reference pass are needed.","The injected noise does not need to be Gaussian, measurable, or independent across nodes: real two-level sMTJ noise, even serially autocorrelated, trains multilayer networks.","Because DANP is local except for the global loss signal, the weight updates and the layer-level decorrelation can in principle be implemented with analog circuitry, reducing the energy cost relative to digital backpropagation.","The approach generalizes beyond magnetic tunnel junctions to any stochastic physical device whose noise can be coupled into node pre-activations, such as memristive devices.","Tunability of sMTJ dwell times, which can reach nanoseconds, sets the bandwidth of the learning process and allows the training loop to speed up as the hardware improves."],"supporting_citations":[{"why":"Introduces the decorrelated activity-based node perturbation (DANP) algorithm that the paper adapts and tests with physical noise.","marker":"[20]"},{"why":"Supplies the iterative decorrelation rule for $R_l$ that upgrades activity-based node perturbation to DANP.","marker":"[30]"},{"why":"Establishes stochastic magnetic tunnel junctions as probabilistic hardware whose two-level fluctuations can be used for computing.","marker":"[16]"},{"why":"Characterizes sMTJ noise and its tunability, providing the physical basis for the measured noise used in training.","marker":"[31]"},{"why":"Gives the Arrhenius dwell-time law used to describe the sMTJ switching timescales.","marker":"[34]"},{"why":"Provides the measured sMTJ switching time series that serves as the real noise input in the simulations.","marker":"[41]"}],"fun_headline_variants":["Physical sMTJ noise teaches networks near-backprop accuracy","Autocorrelated noise from sMTJs learns without backprop","Node-perturbation with real junction noise hits backprop levels","Stochastic junctions: noise as the learning signal","Noise-based learning in magnetic tunnel junctions matches BP"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that injecting the measured sMTJ noise into a software model faithfully represents what would happen in the proposed analog circuit, since the physical experiment itself keeps the network, weights, and updates in software rather than in the resistive network.","fun_headline_variants_meta":{"raw":{"variants":["Physical sMTJ noise teaches networks near-backprop accuracy","Autocorrelated noise from sMTJs learns without backprop","Node-perturbation with real junction noise hits backprop levels","Stochastic junctions: noise as the learning signal","Noise-based learning in magnetic tunnel junctions matches BP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000268,"raw_usage":{"total_tokens":1566,"prompt_tokens":841,"completion_tokens":725,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":457,"completion_tokens_details":{"reasoning_tokens":643}},"tokens_in":457,"tokens_out":725,"duration_ms":7731,"temperature":1.0,"reasoning_tokens":643,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:44:03.129145+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"If a full analog 2x2x2x1 network built from the proposed sMTJ noise-injection circuit is trained and its loss does not decrease over 100 epochs, or if the same loss decrease occurs when the sMTJ is replaced with a fixed voltage, then the claim that physical sMTJ noise drives learning is falsified.","supporting_citations":[{"cited_title":"Ahmad, E","cited_arxiv_id":null,"evidence_quote":"Supplies the iterative decorrelation rule for $R_l$ that upgrades activity-based node perturbation to DANP."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes stochastic magnetic tunnel junctions as probabilistic hardware whose two-level fluctuations can be used for computing."},{"cited_title":"Schnitzspan, M","cited_arxiv_id":null,"evidence_quote":"Characterizes sMTJ noise and its tunability, providing the physical basis for the measured noise used in training."},{"cited_title":"N´ eel, Th´ eorie du tra ˆ ınage magn´ etique des ferromagn´ etiques en grains fins avec application aux terres cuites, in Ann","cited_arxiv_id":null,"evidence_quote":"Gives the Arrhenius dwell-time law used to describe the sMTJ switching timescales."}],"review_version":1}