Pith. sign in

REVIEW 4 major objections 6 minor 87 references

OscNet: Machine Learning on CMOS Oscillator Networks

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

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

Pith's one-line read The paper claims that a CMOS oscillator network can learn MNIST digits with only forward propagation and a Hebbian winner-take-all update, reaching 57.31% unsupervised and 90.58% after linear finetuning.

desk verdict Forward-only Hebbian learning on oscillator networks is promising, but a sign error in the Potts Hamiltonian derivation leaves the core hardware-to-accuracy link unproven. read the letter →

arxiv 2502.07192 v1 pith:OHX3FFY2 submitted 2025-02-11 cs.CV

Wenxiao Cai , Thomas H. Lee This is my paper · ORCID
classification cs.CV
keywords CMOSoscillatornetworksHebbianlearningPottsHamiltonianphaseencodingneuromorphiccomputingunsupervisedMNISTclassificationenergy-efficientAI
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

OscNet proposes to do machine learning in the phases of coupled CMOS oscillators rather than in the voltage levels of a conventional processor. The paper's central claim is that one forward pass through the network minimizes a Potts Hamiltonian, so each output oscillator's phase encodes the circular mean of its weighted inputs. Training uses only a local Hebbian winner-take-all update, with no backward error signal. In simulations this pipeline reaches 57.31% unsupervised accuracy on MNIST and 90.58% after a supervised linear readout, matching or exceeding the autoencoder baselines the paper reports. The payoff would be a hardware fabric that learns and infers in one energy-efficient physical step.

What carries the argument

The load-bearing object is the phase-averaging identity of Eq. 8: an output oscillator's settled phase is the weighted circular mean of its inputs, $\theta_o=\arctan(\sum_i J_i\sin\theta_i/\sum_i J_i\cos\theta_i)$, obtained by setting the derivative of the Potts Hamiltonian to zero. The paper encodes real numbers as tangents of phases, so that reading out an output phase directly yields the normalized weighted sum needed for convolution, regression, clustering, or classification. The Winner-Takes-All Hebbian update of Eq. 17 then moves the winning output's coupling vector toward the current input pattern while decaying all others, which is how the network learns without backpropagation.

What would settle it

Simulate the Kuramoto phase dynamics of Eq. 2 with the exact weight matrices from the MNIST experiments and compare the settled output phases with the circular mean formula of Eq. 8; if any output phase lands away from that formula for a negative coupling weight, the forward-propagation mechanism as stated does not hold for that configuration.

Watch

Extended reading notes

Core claim

The central discovery is that a CMOS oscillator network can function as a trainable feature-extraction layer. With input values encoded as phases $\theta_i=\arctan X_i$ and coupling strengths $J_{ij}$, the paper derives the steady output phase $\theta_{oj}=\arctan\!\left(\sum_i J_{ij}\sin\theta_i\,/\,\sum_i J_{ij}\cos\theta_i\right)$ by minimizing a Potts Hamiltonian, so the network's forward pass is the physical settling of oscillator phases. The same phase-averaging computation is then used for image convolution, K-means clustering, linear regression, and classification. The paper trains the coupling weights with a Winner-Takes-All Hebbian rule and reports the MNIST accuracies as evidence that forward-only, local learning on this hardware model is competitive with autoencoder pipelines.

Load-bearing premise

The results assume that an output oscillator's phase actually converges to the circular mean of Eq. 8 for the mixed positive and negative coupling weights used in the experiments, even though the paper's sign conventions in the energy function are inconsistent.

Editorial extensions

If this is right

  • A single-layer OscNet trained with forward propagation and Hebbian winner-take-all reaches 57.31% unsupervised accuracy on MNIST, and 90.58% after supervised linear finetuning, matching or beating the reported autoencoder baselines at comparable widths.
  • The absence of backpropagation removes the need to store and propagate error gradients, so the learning loop is simpler and more energy-efficient than standard deep learning training.
  • The same phase-averaging forward pass handles convolution, K-means clustering, and linear regression, meaning one oscillator fabric can serve several machine-learning primitives.
  • Because the update rule is local and unsupervised, the network can learn continuously from streaming data in hardware, without a separate offline training stage.

Reading between the lines

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

  • If the phase-averaging model holds at the circuit level, stacking multiple oscillator layers with the same Hebbian rule would be a natural next step, but the paper only demonstrates single-layer networks; testing this on a deeper architecture would show whether the approach scales.
  • The reported accuracies come from numerical simulation of the Kuramoto model, not from fabricated silicon, so the energy-efficiency advantage remains an extrapolation until a physical chip measures phase convergence under noise and process variation.
  • Because the Hamiltonian's sign is handled by fiat ('we ignore the negative sign'), the theoretical convergence guarantee may only be secure for all-positive couplings; separating the MNIST weight matrix into positive-only and mixed-sign runs would test whether the accuracy comes from the claimed physics or from the Hebbian update alone.
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. This paper proposes OscNet, a CMOS oscillator network in which numerical values are encoded as oscillator phases. The authors introduce a MIMO version of OscNet, claim that the free-running output phases minimize a Potts Hamiltonian, and derive the output phase as the circular mean of weighted inputs (Eq. 8). This forward model is then used for image convolution, K-means clustering, linear regression, and Hebbian winner-take-all unsupervised learning. The paper reports a 4-input, 3-output simulation and MNIST experiments: 57.31% unsupervised accuracy (Table 1), 77.47-91.33% after supervised linear finetuning (Table 2), and 60.84% for OscNet K-means (Table 3). The central claim is that forward-propagation-only Hebbian learning on oscillator hardware can match or exceed simple autoencoder baselines while offering an energy-efficient path.

Significance. The idea of using oscillator phase as a computational variable and learning with a local Hebbian rule is potentially attractive for energy-efficient neuromorphic hardware, and the MNIST results are concrete, falsifiable empirical claims. The paper's strengths are its explicit phase-encoding recipes and the direct comparison to autoencoder baselines. However, as written, the theoretical foundation is not sound: the sign inconsistency in the Potts Hamiltonian and the mismatch between Eqs. (15) and (18) mean that the reported experiments are not demonstrably connected to the claimed oscillator dynamics. The significance is therefore contingent on a corrected and fully specified forward model; without that connection, the empirical numbers alone would be a modest contribution.

major comments (4)
  1. [Sec. 2.2, Eqs. (3)-(5) and (8)] The sign conventions are internally inconsistent. Eq. (3) defines a Lyapunov function E(t) with positive cosine terms, while Eq. (4) defines the Potts Hamiltonian H with negative cosine terms; since H = -E up to constants, the dynamics that minimize E actually maximize H. Eq. (5) then states that the output phase is the argmin of Σ J_ij cos(θ - θ_i), but the stationary condition (7) gives the circular mean (8), which for positive J_ij is the argmax, not the argmin. The sentence 'we ignore the negative sign in Eq. 8' is not a legitimate simplification: flipping the sign of an optimization objective exchanges minima and maxima. Concretely, for a single input with J=1 and θ_i=0, Eq. (8) predicts θ=0, while Eq. (5) predicts θ=π. Because both the circuit-level validation in Sec. 6 and the MNIST forward model are justified through Eq. (8), the claim that CMOS oscillator dynamics produce the reported results is not established as written. The authors should either define the output as the argmax of the Potts Hamiltonian or absorb the sign into the coupling strengths, and then state and verify the conditions under which the network converges to that value.
  2. [Sec. 4.1 Eq. (15) and Sec. 4.2 Eq. (18)] The forward pass changes between a sum-normalized weighted average and a plain dot product. Eq. (15) yields θ_j = arctan(Σ W_ij X_i / Σ W_ij), while Eq. (18) yields y_j = Σ x_i w_ij. These are different operations: the former divides by the sum of the weights, the latter does not. Moreover, Eq. (15) is not the direct consequence of Eq. (8) with θ_i = arctan X_i for arbitrary weights; it requires the input-dependent coupling J_ij = W_ij sqrt(1 + X_i^2) used in Sec. 3.1 but not stated in Sec. 4.1. The manuscript does not specify which forward pass was used in the MNIST experiments (Tables 1 and 2) or in the K-means experiment (Table 3). This must be fixed: state the exact forward operation used, derive it from Eq. (8) with the appropriate coupling encoding, and ensure that Secs. 4.1, 4.2, and 6 are mutually consistent.
  3. [Sec. 4.1, Eq. (17) and Algorithm 1] The Hebbian weight update is the learning rule responsible for the MNIST results, but the paper provides no convergence analysis and no empirical study of its behavior. The update is not obviously stable: the winner term W_ij + tanθ_j*(X - tanθ_j* W_ij) and the loser term W_ij - λ tanθ_j*(X - tanθ_j* W_ij) can change magnitude unpredictably for large tanθ_j*, and the choice of λ is not discussed. Algorithm 1 leaves L, the weight initialization scale, and the stopping criterion unspecified. Without these details, the accuracy numbers in Tables 1 and 2 are not reproducible, and there is no evidence that the learned features correspond to a fixed point or minimum of any stated objective. Please add a convergence experiment (e.g., training accuracy or feature evolution versus iteration) and report the hyperparameters, or provide code.
  4. [Sec. 6] The only circuit-level validation is a single 4-input, 3-output simulation. The paper reports the final phases but omits the simulation parameters: the coupling strengths K_ij derived from the W matrix, the pump strength K_p, the integration method and time step, the initial phases, and the stopping criterion. It is therefore impossible to verify that the free-running output phases actually converge to the values in Eq. (8) rather than being artifacts of the chosen initial conditions. Given that the sign inconsistency in Sec. 2.2 directly affects what the dynamics minimize, this validation is load-bearing; it needs full details (or code) and ideally a demonstration of convergence from multiple initial phase conditions.
minor comments (6)
  1. [Sec. 2.2] Typo: 'igonre' appears after Eq. (5), and 'Potts Hamilton' should be 'Potts Hamiltonian'.
  2. [Abstract and Sec. 1] Minor language issues: 'a energy efficient' should be 'an energy-efficient'; 'OscNet is a specially designed hardware' should be 'OscNet is specially designed hardware' or 'OscNet is a specially designed hardware platform'.
  3. [Sec. 2, Eqs. (2)-(4)] The symbol N is used both for the number of oscillator states/injection harmonic order and for the number of oscillators; this is confusing, particularly in Eqs. (2) and (3). Please use a distinct symbol such as N_inj for the harmonic order.
  4. [Sec. 4.1, Eq. (16)] The readout rule j* = argmax_j tanθ_j is stated without explaining how tanθ_j is computed from the oscillator phase θ_j (which is defined modulo 2π). Clarify whether the readout uses the principal value of arctan or the actual phase, and how negative values are handled.
  5. [Sec. 6] The simulation outputs O = [-7.545, -4.600, 28.000] match Σ W_i X_i / Σ W_i for each output column, i.e., the normalized forward pass of Eq. (15). State this explicitly so the reader can see which equation is being validated.
  6. [Sec. 7.2, Table 2] The text says OscNet achieves results 'comparable' to the AutoEncoder; this is true for 10, 16, and 64 hidden neurons, but at 128 hidden neurons the AutoEncoder is slightly better (91.35 vs 91.33). The wording should be adjusted to reflect this.

Circularity Check

2 steps flagged · score 6.0 of 10

The forward-pass identity (Eq. 8) is forced by dropping the Hamiltonian's sign: the circular mean is the argmax of the written Hamiltonian, not its argmin, so the derivation is self-definitional; the convolution and regression 'derivations' inherit that tautology, while the MNIST results are the only independent, non-circular evidence.

  1. self definitional [Section 2.2, Eqs. 5-8]
    "Since Jij can be either positive or negative, for simplicity, we ignore the negative sign in Eq. 8. hi is a time-independent term and can also be ignored in the differentiation. Consider an output oscillator theta_o, its value is given by Potts Hamilton: theta_oj = argmin_theta H = argmin_theta sum_i J_ij cos(theta - theta_i). (5) ... So theta_oj is: theta_oj = theta_0 = arctan( sum_i J_ij sin theta_i / sum_i J_ij cos theta_i ). (8)"

    With positive J, sum_i J_ij cos(theta - theta_i) is maximized when theta equals the circular mean theta_0 (the phase vectors align), and minimized at theta_0 + pi. Therefore theta_0 is not the argmin of the Hamiltonian written in Eq. 5; it is the argmax. The only way Eq. 8 follows from Eq. 5 is to flip the sign of the cosine sum, which is precisely the sign the text says it is ignoring. The sign flip makes the desired circular-mean formula the stationary point by construction, so Eq. 8 is an assumption encoded in the Hamiltonian convention, not a prediction from oscillator dynamics. Convolution (Sec. 3.1) and regression (Sec.

  2. self citation load bearing [Section 2.1]
    "CMOS Oscillator Network [71–73] operates with high-order injection locking, with local connections to neighboring oscillators. ... The CMOS oscillator network can find the minimal value of Potts Hamiltonian and thus minimize energy of the system."

    The hardware property that the network minimizes the Potts Hamiltonian is not established in this paper; it is imported from refs [71-73], authored by R.L. Smith and T.H. Lee, with Lee a co-author of the present work. The current paper does not build or measure the CMOS circuit; Section 6 simulates the Kuramoto equation directly. Thus the central physical premise is justified by a self-citation chain rather than by independent evidence. This is load-bearing because the MNIST and convolution claims all presuppose that the oscillator phases converge to Eq. 8, which the sign-error above shows is not a consequence of the quoted Hamiltonian.

full rationale

The paper's central theoretical step is Eq. 8, the claim that a free-running oscillator phase equals the circular mean of its weighted inputs. As written, Eq. 5 defines the output as the argmin of a positive-cosine sum, but the circular mean theta_0 is the argmax of that sum for positive couplings. The text says it 'ignore[s] the negative sign' and then obtains theta_0 as the stationary solution; this is a sign convention chosen to force the desired result, making Eq. 8 self-definitional rather than derived from the oscillator Hamiltonian. The convolution encoding (Eqs. 10-13) and regression encoding (Eqs. 23-26) are then constructed by solving for the phase and coupling values that make Eq. 8 reproduce the target formulas, so those 'derivations' are consistency checks, not independent physical predictions. The MNIST experiments are genuinely empirical and non-circular; they compare OscNet's Hebbian pipeline against autoencoder baselines and show competitive accuracy. However, those experiments use Eq. 15 as the forward model, which is not the same as the theoretical Eq. 8, and they inherit the unsupported physical premise. The self-citations to the authors' earlier CMOS oscillator work are load-bearing for the hardware claim but are not independently verified here. Overall, the paper contains partial circularity: the core forward-pass identity reduces to a sign convention and the derived applications are tautological, while the empirical benchmarks provide independent but narrower support.

Assumptions & free parameters 5 free parameters · 5 assumptions · 1 invented entities

The central claims rest on the oscillator dynamics converging to a Potts Hamiltonian minimum, the continuous-phase simplification, the ad hoc Hebbian update, and several unreported parameters (lambda, L, initialization, simulation settings). These are the quantities a reader must take on faith; the paper provides no code, error bars, or hardware measurements to constrain them.

free parameters (5)
  • lambda (weight decay coefficient, Eq. 17) = not reported
    Controls how quickly non-winning connections decay in the Hebbian update; the paper gives no value or tuning method, so the reported MNIST accuracies are not reproducible from the text.
  • L (number of training passes, Algorithm 1) = not reported
    The outer loop in Algorithm 1 iterates L times over training data; L is never specified for the MNIST experiments.
  • weight initialization scale = not reported
    Random initial connection weights are required in Algorithm 1; no distribution or scale is given, and results can depend on it.
  • pump strength Kp and simulation parameters (Sec. 6) = not reported
    The Kuramoto simulation in Sec. 6 does not report Kp, integration time, initial phases, or solver settings; the single 4x3 example cannot be reproduced.
  • linear finetuning hyperparameters = not reported
    Supervised finetuning results in Table 2 depend on the linear regression optimizer, regularization, and epochs, none of which are given.
assumptions (5)
  • domain assumption Kuramoto phase dynamics with injection locking (Eq. 2)
    Oscillator interactions are assumed to follow Kuramoto's equation with a master pump, taken from prior OscNet work [71-73].
  • standard math Existence of a Lyapunov function that decreases over time (Eq. 3)
    Used to claim the network minimizes energy; standard for Kuramoto-type systems, cited from [33, 72, 80].
  • ad hoc to paper Infinite N harmonic so phases take continuous values in [0, 2pi) (Sec. 2.2)
    The paper assumes N is infinite to allow continuous phases and arbitrary real tan values; fabricated CMOS networks would have finite N and discrete phases, which could change the outputs.
  • domain assumption The oscillator network minimizes the Potts Hamiltonian of Eq. 5
    The paper asserts global minimization despite mixed-sign couplings; no proof or condition on coupling signs is given.
  • ad hoc to paper Winner-take-all Hebbian updates (Eq. 17) converge to useful features
    No convergence analysis is provided; the MNIST results are the only evidence, and hyperparameters are unreported.
invented entities (1)
  • MIMO OscNet architecture
    purpose: A network with fixed input oscillators and mutually non-interacting output oscillators to compute multiple weighted averages in parallel
    Proposed in this paper; only MATLAB simulations are shown, no measured chip or fabricated prototype, so there is no independent evidence the hardware behaves as modeled.

how reviews work

0 comments
Cite this review

Pith. "Pith review of OscNet: Machine Learning on CMOS Oscillator Networks." pith.science (2026). https://pith.science/paper/OHX3FFY2

@misc{pith2026250207192,
  author       = {Pith},
  title        = {Pith review of: OscNet: Machine Learning on CMOS Oscillator Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OHX3FFY2}},
  note         = {Machine review of arXiv:2502.07192}
}
read the original abstract

Machine learning and AI have achieved remarkable advancements but at the cost of significant computational resources and energy consumption. This has created an urgent need for a novel, energy-efficient computational fabric to replace the current computing pipeline. Recently, a promising approach has emerged by mimicking spiking neurons in the brain and leveraging oscillators on CMOS for direct computation. In this context, we propose a new and energy efficient machine learning framework implemented on CMOS Oscillator Networks (OscNet). We model the developmental processes of the prenatal brain's visual system using OscNet, updating weights based on the biologically inspired Hebbian rule. This same pipeline is then directly applied to standard machine learning tasks. OscNet is a specially designed hardware and is inherently energy-efficient. Its reliance on forward propagation alone for training further enhances its energy efficiency while maintaining biological plausibility. Simulation validates our designs of OscNet architectures. Experimental results demonstrate that Hebbian learning pipeline on OscNet achieves performance comparable to or even surpassing traditional machine learning algorithms, highlighting its potential as a energy efficient and effective computational paradigm.

Figures

Figures reproduced from arXiv: 2502.07192 by the authors.

Figure 1
Figure 1. The Oscillator Network (OscNet) is a CMOS circuit inspired by the brain, where current serves as the medium for information [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Modeling of human visual perception. Light from the real world travels to the retina, where it is processed by retina cells. These [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. MIMO OscNet simulation for 4 inputs and 3 outputs. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Learned features of OscNet and Auto Encoder on [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

87 extracted references · 72 canonical work pages

  1. [1]

    Journal of theoretical Biology 195(4), 419–438 (1998) 8

    Adams, P.: Hebb and darwin. Journal of theoretical Biology 195(4), 419–438 (1998) 8

    1998

  2. [2]

    In: Proceedings of the 2018 ACM international conference on bioinformat- ics, computational biology, and health informatics

    Ahmad, M.A., Eckert, C., Teredesai, A.: Interpretable machine learning in healthcare. In: Proceedings of the 2018 ACM international conference on bioinformat- ics, computational biology, and health informatics. pp. 559–560 (2018) 1

    2018

  3. [3]

    Research in Interna- tional Business and Finance 61, 101646 (2022) 1

    Ahmed, S., Alshater, M.M., El Ammari, A., Ham- mami, H.: Artificial intelligence and machine learning in finance: A bibliometric review. Research in Interna- tional Business and Finance 61, 101646 (2022) 1

    2022

  4. [4]

    PLoS Computational Biology 4 (2008), https : //api.semanticscholar.org/CorpusID: 1493332 4, 8

    Albert, M.V ., Schnabel, A., Field, D.J.: Innate vi- sual learning through spontaneous activity patterns. PLoS Computational Biology 4 (2008), https : //api.semanticscholar.org/CorpusID: 1493332 4, 8

    2008

  5. [5]

    In: Proceedings of the second workshop on simple and efficient natural language processing

    Bannour, N., Ghannay, S., N ´ev´eol, A., Ligozat, A.L.: Evaluating the carbon footprint of nlp methods: a sur- vey and analysis of existing tools. In: Proceedings of the second workshop on simple and efficient natural language processing. pp. 11–21 (2021) 1

    2021

  6. [6]

    Expert Systems with Applications 175, 114820 (2021) 1

    Bertolini, M., Mezzogori, D., Neroni, M., Zammori, F.: Machine learning for industrial applications: A comprehensive literature review. Expert Systems with Applications 175, 114820 (2021) 1

    2021

  7. [7]

    Neural Computation 32(11), 2279– 2309 (2020) 8

    Boutin, V ., Franciosini, A., Ruffier, F., Perrinet, L.: Effect of top-down connections in hierarchical sparse coding. Neural Computation 32(11), 2279– 2309 (2020) 8

    2020

  8. [8]

    Annual review of neuroscience13(1), 475–511 (1990) 8

    Brown, T.H., Kairiss, E.W., Keenan, C.L.: Hebbian synapses: biophysical mechanisms and algorithms. Annual review of neuroscience13(1), 475–511 (1990) 8

    work page 1990

Show all 87 references
  1. [9]

    Science 304, 1926 – 1929 (2004), https : / / api

    Buzs ´aki, G., Draguhn, A.: Neuronal oscillations in cortical networks. Science 304, 1926 – 1929 (2004), https : / / api . semanticscholar . org/CorpusID:8002293 2

    2004

  2. [10]

    arXiv preprint arXiv:2412.18703 (2024) 1

    Cai, W., Hu, D., Yin, R., Deng, J., Fu, H., Yang, W., Gong, M.: Uncertainty quantification in stereo match- ing. arXiv preprint arXiv:2412.18703 (2024) 1

    2024 arXiv

  3. [11]

    arXiv preprint arXiv:2305.13608 (2023)

    Cai, W., Jin, K., Hou, J., Guo, C., Wu, L., Yang, W.: Vdd: Varied drone dataset for semantic segmentation. arXiv preprint arXiv:2305.13608 (2023)

    2023 arXiv

  4. [12]

    arXiv preprint arXiv:2406.13642 (2024)

    Cai, W., Ponomarenko, I., Yuan, J., Li, X., Yang, W., Dong, H., Zhao, B.: Spatialbot: Precise spatial under- standing with vision language models. arXiv preprint arXiv:2406.13642 (2024)

    2024 arXiv

  5. [13]

    arXiv preprint arXiv:2402.12677 (2024) 1

    Cai, W., Yang, W.: Object-level geometric structure preserving for natural image stitching. arXiv preprint arXiv:2402.12677 (2024) 1

    2024 arXiv

  6. [14]

    In: Journal of Neuroscience (1993), https : / / api

    Chapman, B., Stryker, M.P.: Development of ori- entation selectivity in ferret visual cortex and ef- fects of deprivation. In: Journal of Neuroscience (1993), https : / / api . semanticscholar . org/CorpusID:1562405 3

    1993

  7. [15]

    In: Pro- ceedings of the 2nd workshop on sustainable computer systems

    Chien, A.A., Lin, L., Nguyen, H., Rao, V ., Sharma, T., Wijayawardana, R.: Reducing the carbon impact of generative ai inference (today and in 2035). In: Pro- ceedings of the 2nd workshop on sustainable computer systems. pp. 1–7 (2023) 1

    2023

  8. [16]

    Choudhary, S., Sloan, S., Fok, S., Neckar, A., Traut- mann, E., Gao, P., Stewart, T., Eliasmith, C., Boahen, K.: Silicon neurons that compute. In: Artificial Neu- ral Networks and Machine Learning–ICANN 2012: 22nd International Conference on Artificial Neural Networks, Lausann...

    2012

  9. [17]

    The American Mathematical Monthly 94(10), 937– 959 (1987) 8

    Cipra, B.A.: An introduction to the ising model. The American Mathematical Monthly 94(10), 937– 959 (1987) 8

    1987

  10. [18]

    SIAM News 33(6), 1–3 (2000) 8

    Cipra, B.A.: The ising model is np-complete. SIAM News 33(6), 1–3 (2000) 8

    2000

  11. [19]

    Machine Learning 20, 273–297 (1995), https : //api.semanticscholar.org/CorpusID: 52874011 6

    Cortes, C., Vapnik, V .N.: Support-vector networks. Machine Learning 20, 273–297 (1995), https : //api.semanticscholar.org/CorpusID: 52874011 6

    1995

  12. [20]

    PLoS computational biology 10(5), e1003564 (2014) 4, 8

    D ¨ahne, S., Wilbert, N., Wiskott, L.: Slow feature anal- ysis on retinal waves leads to v1 complex cells. PLoS computational biology 10(5), e1003564 (2014) 4, 8

    2014

  13. [21]

    arXiv preprint arXiv:2109.05472 (2021) 1

    Desislavov, R., Mart ´ınez-Plumed, F., Hern ´andez- Orallo, J.: Compute and energy consumption trends in deep learning inference. arXiv preprint arXiv:2109.05472 (2021) 1

    2021 arXiv

  14. [22]

    Dixon, M.F., Halperin, I., Bilokon, P.: Machine learn- ing in finance, vol. 1170. Springer (2020) 1

    2020

  15. [23]

    Erhan, D., Courville, A., Bengio, Y ., Vincent, P.: Why does unsupervised pre-training help deep learning? In: Proceedings of the thirteenth international conference on artificial intelligence and statistics. pp. 201–208. JMLR Workshop and Conference Proceedings (2010) 5

    2010

  16. [24]

    In: Synapse development and maturation, pp

    Feller, M., Kerschensteiner, D.: Retinal waves and their role in visual system development. In: Synapse development and maturation, pp. 367–382. Elsevier (2020) 4, 8

    2020

  17. [25]

    Cell calcium 37(5), 425–432 (2005)

    Firth, S.I., Wang, C.T., Feller, M.B.: Retinal waves: mechanisms and function in visual system develop- ment. Cell calcium 37(5), 425–432 (2005)

    2005

  18. [26]

    Science 373(6553), eabd0830 (2021) 4, 8

    Ge, X., Zhang, K., Gribizis, A., Hamodi, A.S., Sabino, A.M., Crair, M.C.: Retinal waves prime visual mo- tion detection by simulating future optic flow. Science 373(6553), eabd0830 (2021) 4, 8

    2021

  19. [27]

    Biological cybernetics 87(5), 404–415 (2002) 8

    Gerstner, W., Kistler, W.M.: Mathematical formu- lations of hebbian learning. Biological cybernetics 87(5), 404–415 (2002) 8

    2002

  20. [28]

    Journal of Behavioral and Experimental Finance 32, 100577 (2021) 1

    Goodell, J.W., Kumar, S., Lim, W.M., Pattnaik, D.: Artificial intelligence and machine learning in finance: Identifying foundations, themes, and research clusters from bibliometric analysis. Journal of Behavioral and Experimental Finance 32, 100577 (2021) 1

    2021

  21. [29]

    parallel development and coding of neural feature detectors

    Grossberg, S.: Adaptive pattern classification and uni- versal recoding: I. parallel development and coding of neural feature detectors. Biological cybernetics 23(3), 121–134 (1976) 8

    1976

  22. [30]

    In: ICASSP 2021- 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    Gupta, M., Ambikapathi, A., Ramasamy, S.: Hebb- net: A simplified hebbian learning framework to do biologically plausible learning. In: ICASSP 2021- 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 3115–

    2021

  23. [31]

    Current genomics 22(4), 291 (2021) 1

    Habehh, H., Gohel, S.: Machine learning in health- care. Current genomics 22(4), 291 (2021) 1

    2021

  24. [32]

    Psychology press (2005) 5

    Hebb, D.O.: The organization of behavior: A neu- ropsychological theory. Psychology press (2005) 5

    2005

  25. [33]

    Journal of Statistical Physics 72, 145–166 (1993), https : / / api

    van Hemmen, J.L., Wreszinski, W.F.: Lyapunov func- tion for the kuramoto model of nonlinearly coupled oscillators. Journal of Statistical Physics 72, 145–166 (1993), https : / / api . semanticscholar . org/CorpusID:122867392 2

    1993

  26. [34]

    IEEE Journal of Solid- State Circuits 54, 2109–2121 (2019), https : //api.semanticscholar.org/CorpusID: 198356617 2

    Hong, B., Hajimiri, A.: A general theory of in- jection locking and pulling in electrical oscilla- tors—part i: Time-synchronous modeling and in- jection waveform design. IEEE Journal of Solid- State Circuits 54, 2109–2121 (2019), https : //api.semanticscholar.org/CorpusID: 19...

    2019

  27. [35]

    IEEE Transactions on Neural Networks 11(3), 734–738 (2000) 8

    Hoppensteadt, F.C., Izhikevich, E.M.: Pattern recog- nition via synchronization in phase-locked loop neu- ral networks. IEEE Transactions on Neural Networks 11(3), 734–738 (2000) 8

    2000

  28. [36]

    Neurocomputing 134, 198–205 (2014) 5

    Hu, X., Zhang, J., Qi, P., Zhang, B.: Modeling re- sponse properties of v2 neurons using a hierarchical k- means model. Neurocomputing 134, 198–205 (2014) 5

    2014

  29. [37]

    Neural computation 24(9), 2422– 2433 (2012) 8

    Hunt, J.J., Ibbotson, M., Goodhill, G.J.: Sparse coding on the spot: Spontaneous retinal waves suffice for ori- entation selectivity. Neural computation 24(9), 2422– 2433 (2012) 8

    2012

  30. [38]

    Science 354(6312), 603–606 (2016) 8

    Inagaki, T., Haribara, Y ., Igarashi, K., Sonobe, T., Ta- mate, S., Honjo, T., Marandi, A., McMahon, P.L., Umeki, T., Enbutsu, K., et al.: A coherent ising ma- chine for 2000-node optimization problems. Science 354(6312), 603–606 (2016) 8

    2016

  31. [39]

    Inc., T.M.: Matlab version: 9.13.0 (r2022b) (2022), https://www.mathworks.com 6

    2022

  32. [40]

    Izhikevich, E.M., Hoppensteadt, F.C.: Poly- chronous wavefront computations. Int. J. Bi- furc. Chaos 19, 1733–1739 (2009), https : //api.semanticscholar.org/CorpusID: 18305391 2

    2009

  33. [41]

    arXiv preprint arXiv:2209.11883 (2022) 8

    Journ ´e, A., Rodriguez, H.G., Guo, Q., Moraitis, T.: Hebbian deep learning without feedback. arXiv preprint arXiv:2209.11883 (2022) 8

    2022 arXiv

  34. [42]

    Physics Letters A 119(3), 140–144 (1986) 8

    Kazakov, V .A.: Ising model on a dynamical pla- nar random lattice: exact solution. Physics Letters A 119(3), 140–144 (1986) 8

    1986

  35. [43]

    Journal of Neuroscience 40(34), 6584–6599 (2020) 4, 8

    Kim, J., Song, M., Jang, J., Paik, S.B.: Spontaneous retinal waves can generate long-range horizontal con- nectivity in visual cortex. Journal of Neuroscience 40(34), 6584–6599 (2020) 4, 8

    2020

  36. [44]

    Proceedings of the National Academy of Sciences 116(16), 7723–7731 (2019) 8

    Krotov, D., Hopfield, J.J.: Unsupervised learning by competing hidden units. Proceedings of the National Academy of Sciences 116(16), 7723–7731 (2019) 8

    2019

  37. [45]

    arXiv preprint arXiv:2012.12229 (2020) 8

    Lagani, G., Amato, G., Falchi, F., Gennaro, C.: Training convolutional neural networks with heb- bian principal component analysis. arXiv preprint arXiv:2012.12229 (2020) 8

    2020 arXiv

  38. [46]

    Physical Review X 13(3), 031020 (2023) 8

    Lopez-Pastor, V ., Marquardt, F.: Self-learning ma- chines based on hamiltonian echo backpropagation. Physical Review X 13(3), 031020 (2023) 8

    2023

  39. [47]

    Frontiers in physics 2, 5 (2014) 8

    Lucas, A.: Ising formulations of many np problems. Frontiers in physics 2, 5 (2014) 8

    2014

  40. [48]

    Luccioni, S., Jernite, Y ., Strubell, E.: Power hungry processing: Watts driving the cost of ai deployment? In: The 2024 ACM Conference on Fairness, Account- ability, and Transparency. pp. 85–99 (2024) 1

    2024

  41. [49]

    IEEE Transactions on Circuits and Sys- tems I: Regular Papers 62(6), 1591–1598 (2015) 8

    Maffezzoni, P., Bahr, B., Zhang, Z., Daniel, L.: Oscil- lator array models for associative memory and pattern recognition. IEEE Transactions on Circuits and Sys- tems I: Regular Papers 62(6), 1591–1598 (2015) 8

    2015

  42. [50]

    2021 17th International Workshop on Cellular Nanoscale Networks and their Ap- plications (CNNA) pp

    Manganaro, G.: Another look at cellular neu- ral networks. 2021 17th International Workshop on Cellular Nanoscale Networks and their Ap- plications (CNNA) pp. 1–4 (2021), https : //api.semanticscholar.org/CorpusID: 244663050 3

    2021

  43. [51]

    Mead, C., Ismail, M.: Analog VLSI implementation of neural systems, vol. 80. Springer Science & Busi- ness Media (2012) 1

    2012

  44. [52]

    arXiv preprint arXiv:2107.01729 (2021) 8

    Miconi, T.: Hebbian learning with gradients: Hebbian convolutional neural networks with mod- ern deep learning frameworks. arXiv preprint arXiv:2107.01729 (2021) 8

    2021 arXiv

  45. [53]

    Nature Reviews Physics 4(6), 363–379 (2022) 8

    Mohseni, N., McMahon, P.L., Byrnes, T.: Ising ma- chines as hardware solvers of combinatorial optimiza- tion problems. Nature Reviews Physics 4(6), 363–379 (2022) 8

    2022

  46. [54]

    Neuromorphic Computing and Engineering 2(4), 044017 (2022) 8

    Moraitis, T., Toichkin, D., Journ ´e, A., Chua, Y ., Guo, Q.: Softhebb: Bayesian inference in unsupervised hebbian soft winner-take-all networks. Neuromorphic Computing and Engineering 2(4), 044017 (2022) 8

    2022

  47. [55]

    Developmental science 7(2), 141–148 (2004) 8

    Munakata, Y ., Pfaffly, J.: Hebbian learning and devel- opment. Developmental science 7(2), 141–148 (2004) 8

    2004

  48. [56]

    Theory of Probability & Its Applications 9(1), 141–142 (1964) 6

    Nadaraya, E.A.: On estimating regression. Theory of Probability & Its Applications 9(1), 141–142 (1964) 6

    1964

  49. [57]

    Neal, R.M.: Pattern recognition and machine learn- ing, by christopher m. bishop. Technometrics 49 (2007), https : / / api . semanticscholar . org/CorpusID:262612365 6

    2007

  50. [58]

    Neural Networks 2, 459–473 (1989) 8

    Network, F.N.: Optimal unsupervised learning in a single-layer linear. Neural Networks 2, 459–473 (1989) 8

    1989

  51. [59]

    CS294A Lecture notes 72(2011), 1–19 (2011) 5, 7

    Ng, A., et al.: Sparse autoencoder. CS294A Lecture notes 72(2011), 1–19 (2011) 5, 7

    2011

  52. [60]

    arXiv preprint arXiv:1412.6597 (2014) 5

    Paine, T.L., Khorrami, P., Han, W., Huang, T.S.: An analysis of unsupervised pre-training in light of recent advances. arXiv preprint arXiv:1412.6597 (2014) 5

    2014 arXiv

  53. [61]

    In: Artificial Life Conference Proceedings

    Pugh, J., Soltoggio, A., Stanley, K.: Real-time heb- bian learning from autoencoder features for control tasks. In: Artificial Life Conference Proceedings. pp. 202–209. MIT Press One Rogers Street, Cambridge, MA 02142-1209, USA journals-info . . . (2014) 8

    2014

  54. [62]

    In: Neural Information Processing Systems (2019), https : //api.semanticscholar.org/CorpusID: 174798065 3, 8

    Raghavan, G., Thomson, M.: Neural networks grown and self-organized by noise. In: Neural Information Processing Systems (2019), https : //api.semanticscholar.org/CorpusID: 174798065 3, 8

    2019

  55. [63]

    Rai, R., Tiwari, M.K., Ivanov, D., Dolgui, A.: Ma- chine learning in manufacturing and industry 4.0 ap- plications (2021) 1

    2021

  56. [64]

    In: Proceedings of the 2014 international sympo- sium on Low power electronics and design

    Ramasubramanian, S.G., Venkatesan, R., Sharad, M., Roy, K., Raghunathan, A.: Spindle: Spintronic deep learning engine for large-scale neuromorphic comput- ing. In: Proceedings of the 2014 international sympo- sium on Low power electronics and design. pp. 15–20 (2014) 8

    2014

  57. [65]

    Progress of Theoretical Physics 77, 1005–1010 (1987), https : / / api

    Sakaguchi, H., Shinomoto, S., Kuramoto, Y .: Lo- cal and grobal self-entrainments in oscillator lat- tices. Progress of Theoretical Physics 77, 1005–1010 (1987), https : / / api . semanticscholar . org/CorpusID:122040965 2, 6

    1987

  58. [66]

    arXiv preprint arXiv:1705.06963 (2017) 8

    Schuman, C.D., Potok, T.E., Patton, R.M., Birdwell, J.D., Dean, M.E., Rose, G.S., Plank, J.S.: A survey of neuromorphic computing and neural networks in hard- ware. arXiv preprint arXiv:1705.06963 (2017) 8

    2017 arXiv

  59. [67]

    The Journal of Anatomy 199(4), 375–383 (2001) 3

    Sernagor, E., Mehta, V .: The role of early neural ac- tivity in the maturation of turtle retinal function. The Journal of Anatomy 199(4), 375–383 (2001) 3

    2001

  60. [68]

    In: 2018 Second international conference on electronics, communica- tion and aerospace technology (ICECA)

    Shailaja, K., Seetharamulu, B., Jabbar, M.: Machine learning in healthcare: A review. In: 2018 Second international conference on electronics, communica- tion and aerospace technology (ICECA). pp. 910–914. IEEE (2018) 1

    2018

  61. [69]

    Shawe-Taylor, J., Cristianini, N.: Kernel meth- ods for pattern analysis (2004), https : //api.semanticscholar.org/CorpusID: 35730151 6

    2004

  62. [70]

    2021 IEEE Interna- tional Midwest Symposium on Circuits and Sys- tems (MWSCAS) pp

    Smith, R.L., Lee, T.H.: Hybrid frequency do- main simulation method to speed-up analysis of injection locked oscillators. 2021 IEEE Interna- tional Midwest Symposium on Circuits and Sys- tems (MWSCAS) pp. 722–726 (2021), https : //api.semanticscholar.org/CorpusID: 237520896 2

    2021

  63. [71]

    2021 IEEE International Symposium on Cir- cuits and Systems (ISCAS) pp

    Smith, R.L., Lee, T.H.: Modeling of injection lock- ing in neurons for neuromorphic and biomedical sys- tems. 2021 IEEE International Symposium on Cir- cuits and Systems (ISCAS) pp. 1–5 (2021), https: //api.semanticscholar.org/CorpusID: 235566110 1, 2

    2021

  64. [72]

    IEEE Open Journal of Cir- cuits and Systems 4, 156–164 (2023), https : //api.semanticscholar.org/CorpusID: 257796022 1, 2

    Smith, R.L., Lee, T.H.: Polychronous oscilla- tory cellular neural networks for solving graph coloring problems. IEEE Open Journal of Cir- cuits and Systems 4, 156–164 (2023), https : //api.semanticscholar.org/CorpusID: 257796022 1, 2

    2023

  65. [73]

    IEEE Transactions on Circuits and Systems II: Express Briefs 71, 4541–4545 (2024), https: //api.semanticscholar.org/CorpusID: 269642781 1, 2

    Smith, R.L., Lee, T.H.: Quantum computing gate emulation using cmos oscillatory cellular neural net- works. IEEE Transactions on Circuits and Systems II: Express Briefs 71, 4541–4545 (2024), https: //api.semanticscholar.org/CorpusID: 269642781 1, 2

    2024

  66. [74]

    Statistics and Computing 14, 199–222 (2004), https : / / api

    Smola, A., Scholkopf, B.: A tutorial on support vec- tor regression. Statistics and Computing 14, 199–222 (2004), https : / / api . semanticscholar . org/CorpusID:15475 6

    2004

  67. [75]

    In: AL- IFE 2016, the Fifteenth International Conference on the Synthesis and Simulation of Living Systems

    Stanley, K., Pugh, J., Bowren, J.: Fully autonomous real-time autoencoder-augmented hebbian learning through the collection of novel experiences. In: AL- IFE 2016, the Fifteenth International Conference on the Synthesis and Simulation of Living Systems. pp. 382–389. MIT Press (2016) 8

    2016

  68. [76]

    arXiv preprint arXiv:1812.05069 (2018) 5, 7

    Tschannen, M., Bachem, O., Lucic, M.: Recent ad- vances in autoencoder-based representation learning. arXiv preprint arXiv:1812.05069 (2018) 5, 7

    2018 arXiv

  69. [77]

    In: Statistics for Engineering and Information Science (2000), https : / / api

    Vapni, V .N.: The nature of statistical learning theory. In: Statistics for Engineering and Information Science (2000), https : / / api . semanticscholar . org/CorpusID:7138354 6

    2000

  70. [78]

    arXiv preprint arXiv:1709.08102 (2017) 8

    Wang, T., Roychowdhury, J.: Oscillator-based ising machine. arXiv preprint arXiv:1709.08102 (2017) 8

    2017 arXiv

  71. [79]

    In: Unconventional Computation and Natural Computation: 18th International Conference, UCNC 2019, Tokyo, Japan, June 3–7, 2019, Proceed- ings 18

    Wang, T., Roychowdhury, J.: Oim: Oscillator-based ising machines for solving combinatorial optimisa- tion problems. In: Unconventional Computation and Natural Computation: 18th International Conference, UCNC 2019, Tokyo, Japan, June 3–7, 2019, Proceed- ings 18. pp. 232–256. Sp...

    2019

  72. [80]

    In: International Conference on Un- conventional Computation and Natural Computation (2019), https : / / api

    Wang, T., Roychowdhury, J.S.: Oim: Oscillator-based ising machines for solving combinatorial optimisa- tion problems. In: International Conference on Un- conventional Computation and Natural Computation (2019), https : / / api . semanticscholar . org/CorpusID:81979522 2

    2019

  73. [81]

    Na- ture Physics 20(9), 1434–1440 (2024) 8

    Wanjura, C.C., Marquardt, F.: Fully nonlinear neuro- morphic computing with linear wave scattering. Na- ture Physics 20(9), 1434–1440 (2024) 8

    2024

  74. [82]

    Sankhy ¯a: The Indian Journal of Statistics, Series A pp

    Watson, G.S.: Smooth regression analysis. Sankhy ¯a: The Indian Journal of Statistics, Series A pp. 359–372 (1964) 6

    1964

  75. [83]

    Journal of Comparative Neurology 158 (1974), https : / / api

    Wiesel, T.N., Hubel, D.H.: Ordered arrangement of orientation columns in monkeys lacking visual experience. Journal of Comparative Neurology 158 (1974), https : / / api . semanticscholar . org/CorpusID:42437417 3

    1974

  76. [84]

    IEEE Transactions on Neu- ral Networks 16, 781–781 (2003), https : //api.semanticscholar.org/CorpusID: 7406938 6

    Williams, C.K.I.: Learning with kernels: Sup- port vector machines, regularization, optimiza- tion, and beyond. IEEE Transactions on Neu- ral Networks 16, 781–781 (2003), https : //api.semanticscholar.org/CorpusID: 7406938 6

    2003

  77. [85]

    Na- ture 601(7894), 549–555 (2022) 8

    Wright, L.G., Onodera, T., Stein, M.M., Wang, T., Schachter, D.T., Hu, Z., McMahon, P.L.: Deep physi- cal neural networks trained with backpropagation. Na- ture 601(7894), 549–555 (2022) 8

    2022

  78. [86]

    Reviews of Mod- ern Physics 54, 235–268 (1982), https : //api.semanticscholar.org/CorpusID: 120281979 2

    Wu, F.Y .: The potts model. Reviews of Mod- ern Physics 54, 235–268 (1982), https : //api.semanticscholar.org/CorpusID: 120281979 2

    1982

  79. [87]

    In: 2018 IEEE international conference on systems, man, and cybernetics (SMC)

    Zhai, J., Zhang, S., Chen, J., He, Q.: Autoencoder and its various variants. In: 2018 IEEE international conference on systems, man, and cybernetics (SMC). pp. 415–419. IEEE (2018) 5, 7

    2018

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.